Next Article in Journal
A Single Center Study of Genes Involved in Synchronous and Metachronous Multiple Early-Stage Gastric Cancers in Japanese Patients with Current or Former Helicobacter pylori Infection
Next Article in Special Issue
BAP1 Mutations and Pleural Mesothelioma: Genetic Insights, Clinical Implications, and Therapeutic Perspectives
Previous Article in Journal
Immune Resistance in Glioblastoma: Understanding the Barriers to ICI and CAR-T Cell Therapy
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Cancers
Volume 17
Issue 3
10.3390/cancers17030463
cancers-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Radiomics-Based Prediction of Treatment Response to TRuC-T Cell Therapy in Patients with Mesothelioma: A Pilot Study

by
Hubert Beaumont
1,*,
Antoine Iannessi
1,
Alexandre Thinnes
1,
Sebastien Jacques
1 and
Alfonso Quintás-Cardama
2
1
Median Technologies, 06560 Valbonne, France
2
TCR2 Therapeutics, Cambridge, MA 02142, USA
*
Author to whom correspondence should be addressed.
Cancers 2025, 17(3), 463; https://doi.org/10.3390/cancers17030463
Submission received: 16 December 2024 / Revised: 24 January 2025 / Accepted: 27 January 2025 / Published: 29 January 2025
(This article belongs to the Special Issue Biomarkers and Targeted Therapy in Malignant Pleural Mesothelioma)
Browse Figures
Review Reports Versions Notes

Simple Summary

T cell receptor fusion constructs (TRuC-Ts) represent a promising next-generation T cell therapy for solid tumors. To enhance their development, improved patient selection is essential. A pilot study was conducted to evaluate the feasibility and performance of a predictive model for treatment responses in mesothelioma patients, leveraging radiomics and machine learning. Radiomics and delta-radiomics (Δradiomics) features from CT scans were analyzed for reproducibility and informativeness, identifying key features for training a random forest classifier. The model achieved an accuracy of 0.75–0.9 in predicting pleural tumor responses, supporting the design of future studies involving 250–400 tumors. This study demonstrated the reproducibility and effectiveness of radiomics/Δradiomics in relation to tumor localization, emphasizing the need for multiple tumor models to create an integrated patient model. These findings provide a foundation for combining tumor-specific models into a unified approach, improving patient selection for TRuC-T therapy in mesothelioma patients.

Abstract

Background/Objectives: T cell receptor fusion constructs (TRuCs), a next generation engineered T cell therapy, hold great promise. To accelerate the clinical development of these therapies, improving patient selection is a crucial pathway forward. Methods: We retrospectively analyzed 23 mesothelioma patients (85 target tumors) treated in a phase 1/2 single arm clinical trial (NCT03907852). Five imaging sites were involved, the settings for the evaluations were Blinded Independent Central Reviews (BICRs) with double reads. The reproducibility of 3416 radiomics and delta-radiomics (Δradiomics) was assessed. The univariate analysis evaluated correlations at the target tumor level with (1) tumor diameter response; (2) tumor volume response, according to the Quantitative Imaging Biomarker Alliance; and (3) the mean standard uptake value (SUV) response, as defined by the positron emission tomography response criteria in solid tumors (PERCISTs). A random forest model predicted the response of the target pleural tumors. Results: Tumor anatomical distribution was 55.3%, 17.6%, 14.1%, and 10.6% in the pleura, lymph nodes, peritoneum, and soft tissues, respectively. Radiomics/Δradiomics reproducibility differed across tumor localizations. Radiomics were more reproducible than Δradiomics. In the univariate analysis, none of the radiomics/Δradiomics correlated with any response criteria. With an accuracy ranging from 0.75 to 0.9, three radiomics/Δradiomics were able to predict the response of target pleural tumors. Pivotal studies will require a sample size of 250 to 400 tumors. Conclusions: The prediction of responding target pleural tumors can be achieved using a machine learning-based radiomics/Δradiomics analysis. Tumor-specific reproducibility and the average values indicated that using tumor models to create an effective patient model would require combining several target tumor models.

1. Introduction

T cell receptor fusion constructs (TRuCs) represent a next generation engineered T cell platform endowed with the ability to target cancer antigens by leveraging the power of the entire T cell receptor (TCR) in an HLA-independent manner [1]. Despite the remarkable success of chimeric antigen receptors (CARs) in several hematological malignancies, multiple studies have failed to replicate similar results among patients with solid tumors, which has hindered their adoption in that setting [2,3]. While novel platforms such as TRuCs hold great promise for the treatment of solid tumors, improving the benefit/risk ratio of these therapies will require an improved upstream selection of patients to improve response rates and minimize treatment-related toxicities.
Several criteria can be considered for evaluating therapeutic responses with imaging techniques. The Response Evaluation Criteria for Solid Tumors (RECISTs) [4] are the most commonly used criteria in oncology clinical trials, although some limitations have been pointed out, particularly their ability to assess certain types of tumors and to evaluate several unique tumor response subgroups outside of a response classification by RECIST, such as a mixed response and an oligometastatic state [5,6]. An alternative method to RECIST, which is based on the assessment of tumor diameter, is that of evaluating tumor volume. Several groups have described the better sensitivity and reproducibility of tumor volume measurements, and the Quantitative Imaging Biomarker Alliance (QIBA) [7] is currently in the final phase of validation for changes in lung tumors volume assessed through computed tomography (CT). On the other hand, positron emission tomography (PET) imaging offers a different insight into therapeutic responses with the quantification of a standard uptake value (SUV) to assess tumor metabolic changes, although it is not exempt from toxicity such as cumulative radiation exposure [8]. So far, none of these different paradigms of responses have clearly outperformed the others. Therefore, analyzing the three of them would provide a wider, more precise, and perhaps complementary, tumor response assessment.
Over the past decade, radiomics, a high-throughput imaging technique, has shown to be effective in detecting and predicting treatment responses [9] as well as in gauging the subtle changes in a tumor microenvironment not adequately addressed by conventional imaging analyses [10].
Radiomics is an innovative approach to medical imaging that identifies patterns that are invisible to the human eye but can be correlated with biological features and clinical outcomes. Radiomics uses high-dimensional data from radiological images to measure basic morphological features (e.g., tumor volume), and the distribution of voxel intensity (e.g., derived from a histogram) to their joint distributions like the Grey Level Co-occurrence Matrix (GLCM), but also a diversity of other indexes derived from the grey-level run length matrix, the neighborhood grey-level difference matrix, or shape-related features. Frequential analysis (e.g., wavelet, Gabor) is being increasingly used nowadays [11]. Furthermore, while radiomics analyzes a static situation, delta-radiomics (Δradiomics) provide an analysis of feature variation over time, i.e., extracting radiomic features from the same region of interest in the same patient at different acquisition time points [12].
The quantitative features extracted by radiomics can be used to predict tumor and patient outcomes at a given time point, whereas Δradiomics assesses feature variation over different acquisition periods, such as before and after therapy [13]. In patients with mesothelioma, PET radiomic features have been shown to be associated with progression-free survival and overall survival rates [14], whereas such an association has not been identified by a standard CT imaging evaluation. To date, no Δradiomics study has been reported on patients with mesothelioma with any imaging method.
The objective of this pilot study was to evaluate whether radiomics or Δradiomics have the potential to predict the response of patients with mesothelioma in the context of a clinical trial where patients were treated with gavocabtagene autoleucel (gavo-cel), a TRuC T cell therapy targeting mesothelin. We evaluated the reproducibility of radiomics and Δradiomics and their correlation with the therapeutic responses of target lesions, as well as the performance of a machine learning algorithm. These different outcomes were used to provide a power analysis and to suggest a classification scheme.

2. Material and Methods

2.1. Study Data

This retrospective study included 23 patients with advanced mesothelin-expressing mesothelioma treated in a phase 1/2, open label, single arm clinical trial (NCT03907852). In this clinical trial, the patients were assessed by CT and PET imaging in a double read Blinded Independent Central Review (BICR) setting. Imaging evaluations were derived from the estimation of the overall response rate (ORR), which included a complete response (CR) and a partial response (PR), according to RECIST v1.1 evaluation criteria. In addition to the regular one-dimensional measurement stipulated in RECIST, volume measurements were performed in all selected target tumors. The same imaging modality and image-acquisition protocols (including the use of an intravenous contrast) were consistently used at all time points for each patient to ensure uniformity in the comparison of lesions.
In addition, the PET data were analyzed according to PERCIST criteria [15] considering the SUV of the hottest tumor at each time point.

2.2. Definition of the Response

At the target tumor level, three different definitions of responses were considered: (1) stable (+25%, −30% change) or a decrease (−30% decrease) in target tumor diameters (RECIST thresholds) with CT imaging; (2) at least a decrease in tumor volume as defined by QIBA [16] with CT imaging; and (3) at least 30% of the hottest mean tumor SUV with PET (PERCIST threshold). The subjective response of non-target lesions was not considered. For each response definition, the ratio between the number of responding tumors and the total number of tumors was computed as the response rate.

2.3. Radiomic Feature Evaluations

The original CT-based tumor volume evaluations were extracted from the iSee platform (Median Technologies, Valbonne, France), then converted to Nifti format. Radiomic features were calculated using LifeX (https://www.lifexsoft.org/ (accessed on 15 December 2024)) freeware [17].
Delta-radiomics were calculated as the net change between the baseline evaluations and the first time point after baseline using the following formula [18]:
NetChange = FeatureTime point 1 − FeatureBaseline
For the reproducibility study, the relative change in response evaluations was calculated using the following formula:
RelativeChange = (FeatureTimepoint1 − FeatureBaseline)/FeatureBaseline

2.4. Study Plan

Data from the original clinical trial were stratified according to patients’ disease and type of tumor (Figure 1).

2.4.1. Reproducibility of Tumor Measurements

Because the original trial featured a double reading setting (i.e., two independent blinded reviewers) for each patient at each time point, a subset of target tumor measurement could be paired. For each radiomic and Δradiomic, the reproducibility was assessed by measuring the paired difference of the measurements.
Previous studies reported that reproducibility can differ according to tumor localization [19,20]. In our study, we hypothesized that reproducibility is homogeneous within each tumor localization, notably between the responding and non-responding tumors. We documented the variability per tumor localization and tested the equivalence of the inter-localization reproducibility.
For each tumor localization, we compared the variability of each radiomic and Δradiomic feature with those reported in the literature [21,22].

2.4.2. Univariate Analysis

We performed multiple comparisons of the mean radiomics values at baseline, week 4 (W4), W8, and W12 by tumor type. For each radiomic and Δradiomic, we performed univariate analyses by testing the equivalence of the mean values for the responding and non-responding tumors (according to diameter, volume, and mean SUV Ground Truth (GT)) at W4, W8, and W12.

2.4.3. Features Selection

We considered an original set of 3416 radiomics, from which feature selection was required [23]. The selection strategy consisted of four steps: (1) a selection of the most reproducible radiomics and Δradiomics determined by a reproducibility (univariate) analysis; (2) a removal of the redundant radiomics and Δradiomics determined by a correlation coefficient threshold; (3) a removal of radiomics and Δradiomics after a process of recursive feature elimination; and (4) a final selection of the number of radiomic and Δradiomic features recommended to avoid data overfitting.

2.4.4. Model Design for Predicting Response to Treatment

A per-tumor model was trained in processing radiomics or Δradiomics independently for each main tumor localization: pleura, lymph nodes, peritoneum, and soft tissues.

2.4.5. Statistics

Statistics were processed using R CRAN software (V. 4.3.3) [24]. Statistical significance levels were two-sided, with p-values < 0.05 and a 95% confidence interval (CI). Because of the small sample size, non-parametric statistics were preferred.
Sunburst display was performed using the “SunburstR” package (V. 2.1.8).
Reproducibility:
We evaluated reproducibility of each radiomic and Δradiomic feature using concordance correlation coefficients (CCCs) as defined by Lin [25]. For several threshold values of concordance, we reported the number of radiomic features higher than those thresholds. The p-value associated with significant non-inferiority values over the thresholds was tested with a bootstrapping CCC (“boot” and “boot.pval” packages) and adjusted for multiple tests using the “Fuzzysim” package. Then, we compared inter-tissue reproducibility by applying an F-Test for the equality of two variances [26] between the two most prevalent types of tumors and by comparing their CCC CIs.
Univariate analysis:
We tested the non-equivalence of the values across tumor localizations for each radiomics/Δradiomics using multiple Kruskal–Wallis tests with a false discovery rate (FDR) correction (Fuzzysim package) [27].
Typical values for each type of tumor were provided with a CI using the “interpretCI” package.
If non-equivalence of the radiomics values across tumor localizations was found, the tumor localizations were studied separately.
We tested whether radiomics/Δradiomics were associated with the responding/non-responding tumors using the Kruskal–Wallis test with an FDR correction (Fuzzysim package).
Cross-radiomics correlations were measured using the Spearman’s rank correlation coefficient (a non-parametric counterpart to Pearson’s correlation). Redundant features were deleted when correlation coefficients were >0.9.
Features selection:
Features reduction was processed by the following:
  • Selecting the reproducible radiomics/Δradiomics based on CCC values (as previously mentioned).
  • Selecting non-redundant radiomics/Δradiomics using a Spearman’s rank correlation coefficient of 0.9. We relied on the clustering [28] package “heatmaply” and the “fmradio” package for display and to perform data reduction after clustering.
  • Running a recursive feature elimination (RFE) based on the random forest algorithm.
  • Avoiding overfitting issues by establishing an acceptable maximal number of predictors in line with the previous pilot studies [20]. The p-to-n ratio was held at 10.
Model design:
Binary responses were classified using the “caret” package in a cross-validation setting with a random forest algorithm [29].
Proportional and corresponding CIs were calculated using the Clopper–Pearson exact CI model from the “PropCI” package, while multiple comparisons of the proportions (e.g., tumor localization) were computed using the Marascuilo test [30].

3. Results

3.1. Imaging Data

The distribution of patients by modality (CT and PET) and center (n = 5) at baseline is depicted in Figure 2.
The number of patients with CT images taken was 23, 22, 18, and 13 at baseline, W4, W8, and W12, respectively, while there were 17, 17, 5, and 9 patients at baseline and the subsequent time points, respectively, with PET images taken.
Figure 3 summarizes in sunburn the distribution of the CT and PET acquisition parameters used by the five imaging centers.

3.2. Anatomical Distribution of Tumors

The anatomical distribution of the target tumors at baseline in the 23 patients is depicted in Figure 4.
The distribution of target tumors by patient is summarized in Table 1. Overall, 39.1% (9/23) of patients had a single tumor localization.

3.3. Criteria of Response—Corresponding Response Rate

For the main tumor localizations, the response rate at each time point is presented by the tumor diameter (responding or stable disease), tumor volume, and mean SUV in Table 2, Table 3 and Table 4, respectively.

3.4. Reproducibility

Double assessments were performed on 21 patients, for whom 67 pairs of tumors were measured at baseline and at follow-ups. Tumors were distributed in the pleura (n = 35), lymph nodes (n = 14), soft tissues (n = 12), and peritoneum (n = 5). Figure 5 shows the number of radiomics and Δradiomics with the highest CCC values by tumor localization.
For the pleura and lymph node tumors, which were the most prevalent tumors, 41.5% of radiomic features had a significantly different variability (p < 0.05, F-Test) and the volume CCCs were 0.36 (95% CI: 0.31; 0.41) and 0.64 (95% CI: 0.30; 0.84), respectively.
For the pleural and lymph node tumors, 39.1% of Δradiomic features had a significantly different variability (p < 0.05, F-Test) and the volume CCCs were 0.57 (95% CI:0.43; 0.68) and 0.18 (95% CI: −0.03; 0.38), respectively.

3.5. Univariate Analysis

3.5.1. Inter-Tumor Differences in Features Values

In total, 16.0% (95% CI: 15.0%; 17.0%) (n = 536) and 10.0% (95% CI: 9.0%; 11.0%) (n = 338) of the radiomic and △radiomic features, respectively, have different values across tumor type.
After correction for multiple testing (FDR, q = 0.05), these proportions were 2.7% (95% CI: 2.1%; 3.2%) (n = 90) and 0% (n = 0), respectively.
At baseline, the tumor volumes were 62.3 (95% CI: 18.8; 105.8) cm3, 17.8 (95% CI: 5.9; 29.8) cm3, 34.5 (95% CI: 12.0; 57.1) cm3, and 20.2 (95% CI: 6.3; 34.0) cm3 in the pleura, lymph node, peritoneum, and soft tissue tumors, respectively. We found no significant differences between tumor volumes (p = 0.6).

3.5.2. Association Between Radiomics and Responses

For each radiomic/Δradiomic and the main tumor localizations, we tested whether the responder/non-responder populations were significantly different. We tested three paradigms of responses: diameter-based, volume-based, and PERCIST, and the results are presented in Table 5, Table 6 and Table 7, respectively. After FDR correction (q = 0.05), no radiomics/Δradiomics were associated with responses in the diameter, volume, or PERCIST.
Number of radiomic features that had a significant difference of means between the responder/non-responder (Kruskal–Wallis test). Some subcategories were not evaluated (NA) because of the limited response rate.

3.6. Feature Selection and Model Design

To avoid data overfitting, the “10 samples per predictor” rule of thumb is commonly used. Therefore, our model should not rely on more than approximately four, two, two, and two predictors for classifying pleura (n= 47), lymph node (n = 15), peritoneum n = 12), and soft tissue (n = 9) tumors, respectively. We selected a maximum of three predictors for classifying the pleural tumors and two for the other tumors.

3.6.1. Preselection

Considering the reproducible radiomic/Δradiomic features (Figure 5a,b), we removed those with an inter-correlation value >0.9 as shown in the sample cluster map in Figure 6.
The number of radiomics/Δradiomics candidates are depicted in Figure 7a,b.

3.6.2. Feature Selection/Models Performances

The responses of the target pleural tumors were predicted based on three radiomics (three wavelets) measured at baseline. The accuracy was 0.9 (95% CI: 0.6; 0.95), the area under the curve (AUC) = 0.88, at W8, and the accuracy was 0.75 (95% CI: 0.34; 0.96), with an AUC = 0.74, at W12.
Peritoneum tumors were predicted based on two radiomics with an accuracy of 0.67 (95% CI: 0.1; 0.99) and an AUC = 0.61.
The other subclasses (of tumor type/visit) could not be evaluated because of the cross-correlation setting that split the dataset, leaving the test set with missing responder or non-responder data.
The responses of the pleural tumors were predicted based on three Δradiomics (one wavelet and two from density percentiles). The accuracy was 0.71 (95% CI: 0.3; 0.9), with an AUC = 0.74, at W8, and 0.8 (95% CI: 0.3; 0.99), with an AUC = 0.68, at W12. The other subclasses (of tumor type/ visit) could not be tested due to limitations in cross-validation.
The predictions of the SUV responses of the pleura tumors were based on two radiomics (2 wavelets); the accuracy was 0.71 (95% CI: 0.3; 09), with an AUC = 0.68 at W4 and 0.75 (95% CI: 0.2; 0.9), with an AUC = 0.7 at W12. Based on two Δradiomics (two from density percentiles), the prediction had an accuracy of 0.75 (95% CI: 0.2; 0.9) and an AUC = 0.75 at W12.

3.7. Power Analysis

Considering the two most reproducible and informative radiomics and Δradiomics drawn from the feature selection, we estimated the sample size needed to reach significancy in setting shrinkage value at 0.9.
Considering the volume-derived responses and three radiomics predictors at W8, with an AUC = 0.8 and a 40% response rate, the minimum sample size would be 369 patients with 148 responses. At W12, with an AUC = 0.7 and a 27% response rate, the minimum sample size would be 303 patients with 82 responses. When considering two Δradiomics predictors with an AUC = 0.7 and a 30% response rate, the minimum sample size would be 323 patients with 97 responses at W8. At W12, with an AUC = 0.8 and a 30% response rate, the minimum sample size would be 323 patients with 97 responses.
Regarding the SUV-derived responses, when considering two radiomics predictors, at W4, with an AUC = 0.8 and a 20% response rate, the minimum sample size would be 246 patients with 50 responses. At W12, with an AUC = 0.7 and a 40% response rate, the minimum sample size would be 369 patients with 148 responses. Using two Δradiomics predictors (CCC = 0.6), with an AUC = 0.75 and a 36% response rate, the minimum sample size would be 355 patients with 128 responses at W12.

4. Discussion

4.1. Staging, Distribution, and Response Biomarkers

This pilot study documented the anatomical distribution of tumors in patients with mesothelioma and showed the imbalance between the responding/non-responding tumors across location and imaging biomarkers, including longest diameter, volume, and mean SUV.
There was a strong imbalance between the responding/non-responding tumors, which significantly limited the value of a diameter-derived imaging assessment paradigm in favor of volume-derived response criteria using CT as the imaging modality.
Importantly, more than half of the target tumors evaluated in patients with mesothelioma were pleural-based, with the lymph nodes, peritoneum, and soft tissues being the second most frequent sites of disease involvement, in line with the reported patterns of metastasis in patients with malignant pleural mesothelioma described by Collins et al. [31]. Collins et al. also concluded that metastatic spread did not appear to have prognostic implications for overall survival. In our study, brain lesions were excluded, and bone lesions were not selected as a “target” in compliance with RECISTv1.1 guideline recommendations. Additionally, a subset of patients had pleural tumors, but they were not selected as a target for a radiomics analysis, partially due to the RECIST selection criteria requiring a “measurable” supra-centimetric lesion.
Radiomics Robustness
In line with previous studies [32,33], we found that the average radiomics/Δradiomics values varied depending on tumor localization. Therefore, from the standpoint of the average value and reproducibility, different sets of radiomics should be considered for tumor applications according to their localization. Consequently, the models aimed at predicting patient responses should integrate a tumor-specific classification and, for the final system to be practical and widely applicable, it must transition from a tumor-centered approach to a patient-centered one. This would require the implementation of hierarchical or multi-level modeling to ensure the system can address the broader complexities of patient care.
For a predictive model, a higher number of reproducible radiomics/Δradiomics allows more flexibility for feature selection, and thus for training. However, in this study, a small proportion of radiomics/Δradiomics were deemed reproducible and the number of reproducible tumor radiomics/Δradiomics were significantly different between the different anatomic locations, with soft tissue and lymph node tumors featuring the highest and lowest number of reproducible radiomics, respectively. Previous studies have shown that the variability of segmentation is a significant factor impacting radiomics reproducibility [34]. Figure 8 illustrates how the complexity of tumor segmentation can impact radiomics variability.
In contrast to Zhao et al. [21], we found that the volume of pleural tumors had a low CCC value, which might be explained by several factors. First, while the data evaluated by Zhao et al. were lung data, the pleural tumors data of this pilot study were probably more complex to segment. Second, the dataset average volume was largely different between the two studies (around 22.0 cm3 and 62.3 cm3 for the lung data and this study, respectively). Even if a lower relative error can be expected for the larger segmentations, these are more likely to be adjacent to other anatomical structures and, therefore, are prone to segmentation errors. In addition, it is worth noting that pleural mesothelioma involvement causes a particular pattern of growth and spread, with circumferential pleural spread being more prevalent than the typical nodular growth observed in most solid tumors. Consequently, the identification of target lesions with a nodular morphology can be challenging in patients with mesothelioma and adds a layer of variability between two independent blinded observers (Figure 8). For this reason, modified RECIST criteria adapted to the evaluation of pleural mesothelioma have been developed [35].

4.2. Radiomics Correlations

While no radiomics or Δradiomics were associated with tumor responses in the univariate analyses, a multivariate classification allowed us to predict the responses of pleural tumors.
Since a comprehensive evaluation of our predictive system was applied to document the distribution of responses for each tumor location (N = 4) and time point (N = 3), the response criteria would ideally feature an acceptable sample size and balance between the responder/non-responder for each subclass. We considered three different criteria for tumor responses, two for CT (tumor diameter and volume) and one for PET (mean SUV). In agreement with Cai et al. [36], we observed that tumor diameter had a different sensitivity/specificity than the volume measurements in CT. The responses derived from the CT diameter measurements were deemed suboptimal for a prediction assessment because of the strong imbalance between the responding/non-responding tumors. Therefore, we focused on the response in volume with CT and on the mean SUV with PET.
In our univariate analysis, some of the 3416 radiomics were correlated to the responses before an FDR correction. After an FDR correction, none of the radiomics or Δradiomics correlated to any response criteria, as previously reported by Chalkidou et al. [37], thereby supporting the use of multivariate models and machine learning.
Our feature selection scheme aimed to select a maximum of two to three radiomics/Δradiomics per tumor localization to avoid data overfitting in a cross-validation setting. Significant classification performances were obtained only for the pleural tumors with an accuracy ranging from 0.9 to 0.75, which could power future studies involving a data sample size of 250 to 400 target tumors.

4.3. Radiomics and Study Limitations

Several study limitations are worth discussing. First, due to the retrospective nature of our study, we had to adapt the measurements from the original trial (NCT03907852), which prevented us from relying on more specific response criteria [35]. Therefore, we applied the standard RECIST 1.1 criteria, and in particular, the volume-derived response criteria, to all mesothelioma tumors regardless of their anatomical location, even though the volume was originally qualified for advanced lung disease. Although QIBA’s profile [16] clearly highlights the limitations of extending the use of volumetry to other diseases, we adopted these criteria because of the high level of qualification.
Second, we grouped all the tumors into four main localizations. The tumors were initially labeled by each independent reader, then grouping was subjectively performed on a radiologic basis by a single third-party radiologist. Aware that this process was prone to inter-group and intra-group errors [38], we attempted to balance the limited sample size and our hypothesis that each tumor localization had a different average value and reproducibility.
Third, avoiding data overfitting was a challenge due to the small sample size of some subclasses. No well-established statistics are currently available for sizing the maximum number of predictors. However, since the univariate analysis showed no correlation, the minimum number of predictors (2–3) was selected to minimize the risk of data overfitting.
Fourth, we were uncertain about the complexity and potential non-linear relationships between the features and target variables. Considering the small sample size, we aimed to minimize the risk of overfitting and the impact of noise on the data. Therefore, we chose Random Forest as a robust and flexible classifier. Although we considered testing other classification methods, comparing algorithm performance was beyond the scope of our study. Fifth, one aspect that was not addressed in this study is how to translate the models designed at the target tumor level into a patient predictive system. We analyzed the correlation and prediction of target tumor responses using the radiomics/Δradiomics values at baseline since these measurements can only be obtained for measurable tumors (target lesions). However, the real value of a predictive system lies in its ability to assist with patient management, which brings up the issue of response criteria. The response evaluation by the RECIST criteria combines the assessment of target, non-target, and new tumors. However, our analysis was performed only on target lesions and, therefore, such analysis would require the use of validated radiomics-derived, target-lesion-only-based response criteria, which is yet to be developed. Such response criteria will need to address the cases where the responding patients have a dissociated response within a specific tumor site and, even more challenging, the cases where dissociated responses occur at different tumor sites.
Lastly, a proper radiomics study is expected to explain the outcome from the imaging and clinical point of view [39]. We did not address this point because it was considered out of the scope of a pilot study [40] and it could have led to premature conclusions.

4.4. Perspectives

The upcoming pivotal study is anticipated to validate the predictive performance of our pilot study. However, the radiomics-only approach presented here is just a foundational step that could be significantly enhanced by incorporating additional imaging biomarkers, such as nutritional biomarkers (e.g., L3—Skeletal Muscle Index) [41], as well as non-imaging biomarkers (e.g., Krebs von den Lungen-6 (KL-6)) [42]. Furthermore, advancements in AI models and the integration of complementary technologies could further elevate the system’s capabilities [43].
Given the ability of radiomics to predict the therapeutic response of target lesions, we could also start considering evaluating the potential of radiomics for clinical trial monitoring, thereby improving the current response criteria used for mesothelioma.

5. Conclusions

This study supports the use of radiomics/Δradiomics and machine learning for response prediction in patients with malignant mesothelioma. Future studies will have to be powered by assessing reproducibility as well as addressing the different response paradigms and their limitations. As the evidence shows that technologies can be efficient, additional resources and imaging data should be implemented in the evaluation of patients with mesothelioma. We believe that radiomics/Δradiomics and machine learning provide the means to address response prediction in patients with mesothelioma, which thus far has not been adequately addressed. In addition to the predictive model described herein, we propose that the analysis of responses per tumor location could also aid in better understanding drug efficacy in the future.

Author Contributions

Conceptualization, A.Q.-C. and H.B.; methodology, H.B.; software, S.J. and A.T.; validation, H.B. and A.I.; formal analysis, H.B. and A.I.; investigation, H.B.; resources, S.J.; data curation, A.T.; writing—original draft preparation, H.B.; writing—review and editing, A.Q.-C., A.I. and H.B.; visualization, H.B.; supervision, A.Q.-C. and S.J.; project administration, S.J. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Institutional Review Board Statement

Our study was approved/waived by the Ethical Committee/Institutional Review Board (IRB) due to its retrospective nature.

Informed Consent Statement

Written informed consent was not required for this study as it did not impact patient management.

Data Availability Statement

The data that support the findings of this study are the property of TCR2 therapeutics. Restrictions apply to the availability of these data, which were used under license for this study.

Acknowledgments

We would like to thank Udayan Guha for his valuable advice and his commitment to this study.

Conflicts of Interest

H.B., A.I., A.T. and S.J. are employees of Median Technologies and A.Q.-C. is employee of TCR2 therapeutics.

Abbreviations

AUCArea Under the Curve
BICRBlinded Independent Central Review
CCCConcordance Correlation Coefficients
CTComputed tomography
FDRFalse Discovery Rate
GLCMGrey Level Co-occurrence Matrix
GTGround Truth
PERCISTPositron Emission Tomography Response Criteria in Solid Tumor
PETPositron Emission Tomograph
ORROverall Response Rate
QIBAQuantitative Imaging Biomarker Alliance
RECISTResponse Evaluation Criteria in Solid Tumors
SUVStandard Uptake Value
TRuC-TT cell receptor fusion construct

References

  1. Baeuerle, P.A.; Ding, J.; Patel, E.; Thorausch, N.; Horton, H.; Gierut, J.; Scarfo, I.; Choudhary, R.; Kiner, O.; Krishnamurthy, J.; et al. Synthetic TRuC receptors engaging the complete T cell receptor for potent anti-tumor response. Nat. Commun. 2019, 10, 2087. [Google Scholar] [CrossRef] [PubMed]
  2. Xiao, B.F.; Zhang, J.T.; Zhu, Y.G.; Cui, X.R.; Lu, Z.M.; Yu, B.T.; Wu, N. Chimeric Antigen Receptor T-Cell Therapy in Lung Cancer: Potential and Challenges. Front. Immunol. 2021, 12, 782775. [Google Scholar] [CrossRef] [PubMed]
  3. Chen, L.; Chen, F.; Li, J.; Pu, Y.; Yang, C.; Wang, Y.; Lei, Y.; Huang, Y. CAR-T cell therapy for lung cancer: Potential and perspective. Thorac. Cancer 2022, 13, 889–899. [Google Scholar] [CrossRef] [PubMed]
  4. Eisenhauer, E.A.; Therasse, P.; Bogaerts, J.; Schwartz, L.H.; Sargent, D.; Ford, R.; Dancey, J.; Arbuck, S.; Gwyther, S.; Mooney, M.; et al. New response evaluation criteria in solid tumours: Revised RECIST guideline (version 1.1). Eur. J. Cancer 2009, 45, 228–247. [Google Scholar] [CrossRef] [PubMed]
  5. Humbert, O.; Chardin, D. Dissociated Response in Metastatic Cancer: An Atypical Pattern Brought Into the Spotlight With Immunotherapy. Front. Oncol. 2020, 10, 566297. [Google Scholar] [CrossRef] [PubMed]
  6. Chen, D.T.; Chan, W.; Thompson, Z.J.; Thapa, R.; Beg, A.A.; Saltos, A.N.; Chiappori, A.A.; Gray, J.E.; Haura, E.B.; Rose, T.A.; et al. Utilization of target lesion heterogeneity for treatment efficacy assessment in late stage lung cancer. PLoS ONE 2021, 16, e0252041. [Google Scholar] [CrossRef] [PubMed]
  7. Buckler, A.J.; Bresolin, L.; Dunnick, N.R.; Sullivan, D.C. A collaborative enterprise for multi-stakeholder participation in the advancement of quantitative imaging. Radiology 2011, 258, 906–914. [Google Scholar] [CrossRef]
  8. Hosono, M.; Takenaka, M.; Monzen, H.; Tamura, M.; Kudo, M.; Nishimura, Y. Cumulative radiation doses from recurrent PET–CT examinations. Br. J. Radiol. 2021, 94, 20210388. [Google Scholar] [CrossRef]
  9. Huemer, F.; Leisch, M.; Geisberger, R.; Melchardt, T.; Rinnerthaler, G.; Zaborsky, N.; Greil, R. Combination strategies for immune-checkpoint blockade and response prediction by artificial intelligence. Int. J. Mol. Sci. 2020, 21, 2856. [Google Scholar] [CrossRef]
  10. Devkota, L.; Starosolski, Z.; Rivas, C.H.; Starosolski, Z.; Annapragada, A.; Ghaghada, K.B.; Parihar, R. Detection of response to tumor microenvironment targeted cellular immunotherapy using nano-radiomics. Sci. Adv. 2020, 6, eaba6156. [Google Scholar] [CrossRef]
  11. Corso, R.; Stefano, A.; Salvaggio, G.; Comelli, A. Shearlet Transform Applied to a Prostate Cancer Radiomics Analysis on MR Images. Mathematics 2024, 12, 1296. [Google Scholar] [CrossRef]
  12. Cousin, F.; Louis, T.; Dheur, S.; Aboubakar, F.; Ghaye, B.; Occhipinti, M.; Vos, W.; Bottari, F.; Paulus, A.; Sibille, A.; et al. Radiomics and Delta-Radiomics Signatures to Predict Response and Survival in Patients with Non-Small-Cell Lung Cancer Treated with Immune Checkpoint Inhibitors. Cancers 2023, 15, 1968. [Google Scholar] [CrossRef] [PubMed]
  13. Nardone, V.; Reginelli, A.; Grassi, R.; Boldrini, L.; Vacca, G.; D’Ippolito, E.; Annunziata, S.; Farchione, A.; Belfiore, M.P.; Desideri, I.; et al. Delta radiomics: A systematic review. Radiol. Med. 2021, 126, 1571–1583. [Google Scholar] [CrossRef] [PubMed]
  14. Pavic, M.; Bogowicz, M.; Kraft, J.; Vuong, D.; Mayinger, M.; Kroeze, S.G.C.; Friess, M.; Frauenfelder, T.; Andratschke, N.; Huellner, M.; et al. FDG PET versus CT radiomics to predict outcome in malignant pleural mesothelioma patients. EJNMMI Res. 2020, 10, 81. [Google Scholar] [CrossRef]
  15. Joo Hyun, O.; Lodge, M.A.; Wahl, R.L. Practical percist: A simplified guide to PET response criteria in solid tumors 1.0. Radiology 2016, 280, 576–584. [Google Scholar] [CrossRef]
  16. Quantitative Imaging Biomarkers Alliance. QIBA Profile : CT Tumor Volume Change for Advanced Disease (CTV-AD). 2016, C, pp. 1–48. Available online: https://qibawiki.rsna.org/images/3/3c/QIBACTVol_TumorVolumeChange_2022Jul21.pdf (accessed on 15 December 2024).
  17. Nioche, C.; Orlhac, F.; Boughdad, S.; Reuze, S.; Goya-Outi, J.; Robert, C.; Pellot-Barakat, C.; Soussan, M.; Frouin, F.E.; Buvat, I. Lifex: A freeware for radiomic feature calculation in multimodality imaging to accelerate advances in the characterization of tumor heterogeneity. Cancer Res. 2018, 78, 4786–4789. [Google Scholar] [CrossRef] [PubMed]
  18. Lin, P.; Yang, P.F.; Chen, S.; Shao, Y.Y.; Xu, L.; Wu, Y.; Teng, W.; Zhou, X.Z.; Li, B.H.; Luo, C.; et al. A Delta-radiomics model for preoperative evaluation of Neoadjuvant chemotherapy response in high-grade osteosarcoma. Cancer Imaging 2020, 20, 7. [Google Scholar] [CrossRef]
  19. Pavic, M.; Bogowicz, M.; Würms, X.; Glatz, S.; Finazzi, T.; Riesterer, O.; Roesch, J.; Rudofsky, L.; Friess, M.; Veit-Haibach, P.; et al. Influence of inter-observer delineation variability on radiomics stability in different tumor sites. Acta Oncol. 2018, 57, 1070–1074. [Google Scholar] [CrossRef] [PubMed]
  20. Park, J.E.; Park, S.Y.; Kim, H.J.; Kim, H.S. Reproducibility and generalizability in radiomics modeling: Possible strategies in radiologic and statistical perspectives. Korean J. Radiol. 2019, 20, 1124–1137. [Google Scholar] [CrossRef]
  21. Zhao, B.; Tan, Y.; Tsai, W.Y.; Qi, J.; Xie, C.; Lu, L.; Schwartz, L.H. Reproducibility of radiomics for deciphering tumor phenotype with imaging. Sci. Rep. 2016, 6, 23428. [Google Scholar] [CrossRef]
  22. van Timmeren, J.E.; Leijenaar, R.T.H.; van Elmpt, W.; Wang, J.; Zhang, Z.; Dekker, A.; Lambin, P. Test–Retest Data for Radiomics Feature Stability Analysis: Generalizable or Study-Specific? Tomography 2016, 2, 361–365. [Google Scholar] [CrossRef] [PubMed]
  23. Demircioğlu, A. Evaluation of the dependence of radiomic features on the machine learning model. Insights Imaging 2022, 13, 28. [Google Scholar] [CrossRef] [PubMed]
  24. Team RDC. R: A Language and Environment for Statistical Computing [Internet]; R Foundation for Statistical Computing: Vienna, Austria, 2011; p. 409. Available online: http://www.r-project.org (accessed on 15 December 2024).
  25. Lin, L.I. A concordance correlation coefficient to evaluate reproducibility. Biometrics 1989, 45, 255–268. [Google Scholar] [CrossRef] [PubMed]
  26. Snedecor, G.W.; Cochran, W.G. Statistical Methods; Iowa State University Press: Ames, IA, USA, 1989. [Google Scholar]
  27. Benjamini, Y.; Hochberg, Y. Controlling the False Discovery Rate: A Practical and Powerful Approach to Multiple Testing. J. R. Stat. Soc. Ser. B 1995, 57, 289–300. [Google Scholar] [CrossRef]
  28. Wilkinson, L.; Friendly, M. History corner the history of the cluster heat map. Am. Stat. 2009, 63, 179–184. [Google Scholar] [CrossRef]
  29. Couronné, R.; Probst, P.; Boulesteix, A.L. Random forest versus logistic regression: A large-scale benchmark experiment. BMC Bioinform. 2018, 19, 270. [Google Scholar] [CrossRef]
  30. Marascuilo, L.A. Extensions of the significance test for one-parameter signal detection hypotheses. Psychometrika 1970, 35, 237–243. [Google Scholar] [CrossRef]
  31. Collins, D.C.; Sundar, R.; Constantinidou, A.; Dolling, D.; Yap, T.A.; Popat, S.; O’Brien, M.E.; Banerji, U.; de Bono, J.S.; Lopez, J.S.; et al. Radiological evaluation of malignant pleural mesothelioma—Defining distant metastatic disease. BMC Cancer 2020, 20, 1210. [Google Scholar] [CrossRef] [PubMed]
  32. Lee, S.H.; Cho, H.H.; Kwon, J.; Lee, H.Y.; Park, H. Are radiomics features universally applicable to different organs? Cancer Imaging 2021, 21, 31. [Google Scholar] [CrossRef] [PubMed]
  33. Beaumont, H.; Iannessi, A.; Bertrand, A.-S.; Cucchi, J.M.; Lucidarme, O. Harmonization of radiomic feature distributions: Impact on classification of hepatic tissue in CT imaging. Eur. Radiol. 2021, 31, 6059–6068. [Google Scholar] [CrossRef] [PubMed]
  34. Haarburger, C.; Müller-Franzes, G.; Weninger, L.; Kuhl, C.; Truhn, D.; Merhof, D. Radiomics feature reproducibility under inter-rater variability in segmentations of CT images. Sci. Rep. 2020, 10, 12688. [Google Scholar] [CrossRef]
  35. Tsao, A.S.; Gladish, G.W.; Gill, R.R. Revised Modified RECIST Criteria in Malignant Pleural Mesothelioma (Version 1.1): A Step Forward in a Long Race. J. Thorac. Oncol. 2018, 13, 871–873. [Google Scholar] [CrossRef]
  36. Cai, W.-L.; Hong, G.-B. Quantitative image analysis for evaluation of tumor response in clinical oncology. Chronic Dis. Transl. Med. 2018, 4, 18–28. [Google Scholar] [CrossRef] [PubMed]
  37. Chalkidou, A.; O’Doherty, M.J.; Marsden, P.K. False discovery rates in PET and CT studies with texture features: A systematic review. PLoS ONE 2015, 10, e0124165. [Google Scholar] [CrossRef] [PubMed]
  38. De Paoli, L.; Quaia, E.; Poillucci, G.; Gennari, A.; Cova, M.A. Imaging characteristics of pleural tumours. Insights Imaging 2015, 6, 729–740. [Google Scholar] [CrossRef]
  39. Lambin, P.; Leijenaar, R.T.H.; Deist, T.M.; Peerlings, J.; de Jong, E.E.C.; van Timmeren, J.; Sanduleanu, S.; Larue, R.T.H.M.; Even, A.J.G.; Jochems, A.; et al. Radiomics: The bridge between medical imaging and personalized medicine. Nat. Rev. Clin. Oncol. 2017, 14, 749–762. [Google Scholar] [CrossRef] [PubMed]
  40. Lee, E.C.; Whitehead, A.L.; Jacques, R.M.; Julious, S.A. The statistical interpretation of pilot trials: Should significance thresholds be reconsidered? BMC Med. Res. Methodol. 2014, 14, 41. [Google Scholar] [CrossRef] [PubMed]
  41. Mutlu, E.; Inanc, M. Prognostic significance of inflammation scores in malignant mesothelioma. Eur. Rev. Med. Pharmacol. Sci. 2024, 28, 2340–2350. [Google Scholar] [CrossRef] [PubMed]
  42. Stockhammer, P.; Baumeister, H.; Ploenes, T.; Bonella, F.; Theegarten, D.; Dome, B.; Pirker, C.; Berger, W.; Hegedüs, L.; Baranyi, M.; et al. Krebs von den Lungen 6 (KL-6) is a novel diagnostic and prognostic biomarker in pleural mesothelioma. Lung Cancer 2023, 185, 107360. [Google Scholar] [CrossRef]
  43. Maniaci, A.; Fakhry, N.; Chiesa-Estomba, C.; Lechien, J.R.; Lavalle, S. Synergizing ChatGPT and general AI for enhanced medical diagnostic processes in head and neck imaging. Eur. Arch. Oto-Rhino-Laryngol. 2024, 281, 3297–3298. [Google Scholar] [CrossRef] [PubMed]
Cancers 17 00463 g001
Figure 1. Analysis workflow. From the original clinical trial, only mesothelioma patients were considered. This study consisted of a data analysis (blue), radiomics and Δradiomics analysis (green), and model design (orange). The outcome was a pilot evaluation of the predictive performances of target tumors responses.
Figure 1. Analysis workflow. From the original clinical trial, only mesothelioma patients were considered. This study consisted of a data analysis (blue), radiomics and Δradiomics analysis (green), and model design (orange). The outcome was a pilot evaluation of the predictive performances of target tumors responses.
Cancers 17 00463 g001
Cancers 17 00463 g002
Figure 2. Distribution of patients by modality and center at baseline. Five imaging centers participated in the evaluation of patients treated in this study. CT and PET imaging were performed in 23 and 17 patients, respectively. A single imaging center performed only CT (Center #1).
Figure 2. Distribution of patients by modality and center at baseline. Five imaging centers participated in the evaluation of patients treated in this study. CT and PET imaging were performed in 23 and 17 patients, respectively. A single imaging center performed only CT (Center #1).
Cancers 17 00463 g002
Cancers 17 00463 g003
Figure 3. Distribution of acquisition parameters at baseline. (Left) CT acquisition parameters. From inner to outer circle: manufacturer (Siemens or GE), models, Kvp (90, 120, 130), reconstruction kernel (Standard, Br40, Soft), slice thickness (2.5; 5), and voxel size (0.6; 1.0). (Right) PET acquisition parameters. From inner to outer circle: manufacturers (Siemens (Siemens Healthineers, Forchheim, Germany), GE (GE Healthcare, Milwaukee, WI, US)), models, reconstruction kernel (AllPass, XYZ Gauss (2.0, 3.5, 5.0)), slice thickness (2.0; 5.5), and voxel size (1.5; 6.9). Representativeness was deemed significant for generalization.
Figure 3. Distribution of acquisition parameters at baseline. (Left) CT acquisition parameters. From inner to outer circle: manufacturer (Siemens or GE), models, Kvp (90, 120, 130), reconstruction kernel (Standard, Br40, Soft), slice thickness (2.5; 5), and voxel size (0.6; 1.0). (Right) PET acquisition parameters. From inner to outer circle: manufacturers (Siemens (Siemens Healthineers, Forchheim, Germany), GE (GE Healthcare, Milwaukee, WI, US)), models, reconstruction kernel (AllPass, XYZ Gauss (2.0, 3.5, 5.0)), slice thickness (2.0; 5.5), and voxel size (1.5; 6.9). Representativeness was deemed significant for generalization.
Cancers 17 00463 g003
Cancers 17 00463 g004
Figure 4. Distribution of target tumors in patients’ anatomy. Out of 85 tumors, 55.3% (n = 47) were found in the pleura, 17.6% (n = 15) in the lymph nodes, 14.1% (n = 12) in the peritoneum, and 10.6% (n = 9) in soft tissues. Two additional tumors, one adrenal tumor and one liver tumor, were classified as “miscellaneous”.
Figure 4. Distribution of target tumors in patients’ anatomy. Out of 85 tumors, 55.3% (n = 47) were found in the pleura, 17.6% (n = 15) in the lymph nodes, 14.1% (n = 12) in the peritoneum, and 10.6% (n = 9) in soft tissues. Two additional tumors, one adrenal tumor and one liver tumor, were classified as “miscellaneous”.
Cancers 17 00463 g004
Cancers 17 00463 g005
Figure 5. Radiomic/Δradiomic features according to different CCCs threshold values. The number of radiomics (a) and Δradiomics (b) were calculated for different threshold values of CCC. Radiomics reproducibility depended on tumor localization, with soft tissues (range: 238; 139) and lymph nodes (range: 43; 0) being the most and least reproducible, respectively. The reproducibility of Δradiomics depended on tumor localization, with soft tissues (range: 101; 53) and lymph nodes (range: 3; 0) being the most and least reproducible, respectively.
Figure 5. Radiomic/Δradiomic features according to different CCCs threshold values. The number of radiomics (a) and Δradiomics (b) were calculated for different threshold values of CCC. Radiomics reproducibility depended on tumor localization, with soft tissues (range: 238; 139) and lymph nodes (range: 43; 0) being the most and least reproducible, respectively. The reproducibility of Δradiomics depended on tumor localization, with soft tissues (range: 101; 53) and lymph nodes (range: 3; 0) being the most and least reproducible, respectively.
Cancers 17 00463 g005
Cancers 17 00463 g006
Figure 6. Sample cluster map of reproducible pleural tumors radiomics. Left: correlation matrix of 21 radiomics that were deemed reproducible (CCC > 0.8); some of them were highly inter-correlated (yellow clusters). Right: after removing highly inter-correlated radiomics (correlation > 0.9), 8 reproducible and non-redundant pleura radiomics were preselected.
Figure 6. Sample cluster map of reproducible pleural tumors radiomics. Left: correlation matrix of 21 radiomics that were deemed reproducible (CCC > 0.8); some of them were highly inter-correlated (yellow clusters). Right: after removing highly inter-correlated radiomics (correlation > 0.9), 8 reproducible and non-redundant pleura radiomics were preselected.
Cancers 17 00463 g006
Cancers 17 00463 g007
Figure 7. Number of radiomic/Δradiomic candidates. We considered a different threshold of CCC values ranging from 0.7 to 0.9 for radiomics (a) and 0.6 to 0.9 for Δradiomics (b). For radiomics and Δradiomics, peritoneum and lymph node tumors were the most and least reproducible tumors, respectively.
Figure 7. Number of radiomic/Δradiomic candidates. We considered a different threshold of CCC values ranging from 0.7 to 0.9 for radiomics (a) and 0.6 to 0.9 for Δradiomics (b). For radiomics and Δradiomics, peritoneum and lymph node tumors were the most and least reproducible tumors, respectively.
Cancers 17 00463 g007
Cancers 17 00463 g008
Figure 8. Sample variability of segmentation and radiomics values. One pleura and one soft tissue tumor were segmented by Reader 1 and Reader 2. The volume, the joint entropy, and the sum of variance (computed from GLCM) were derived from the segmentations. The inter-reader variability of the segmentations leads to a variability in volume of 20% and 30%, in joint entropy of 16% and 7%, and in sum of variance of 41% and 56%, respectively, the pleura and the soft tissue tumors.
Figure 8. Sample variability of segmentation and radiomics values. One pleura and one soft tissue tumor were segmented by Reader 1 and Reader 2. The volume, the joint entropy, and the sum of variance (computed from GLCM) were derived from the segmentations. The inter-reader variability of the segmentations leads to a variability in volume of 20% and 30%, in joint entropy of 16% and 7%, and in sum of variance of 41% and 56%, respectively, the pleura and the soft tissue tumors.
Cancers 17 00463 g008
Table 1. Distribution of target tumor localization by patient at baseline.
Table 1. Distribution of target tumor localization by patient at baseline.
Tumor LocalizationDisease in PatientsNumb. of Patient with a Unique Disease
Pleura135
Lymph nodes60
Peritoneum83
Soft tissues41
Miscellaneous20
Rows report the main tumor localizations. Central column: number of patients for whom the corresponding tumor localization was reported. Right column: number of patients for whom a single tumor localization was reported.
Table 2. Response rate derived from tumor diameter. Response rate is presented by percentages and proportions. With a response rate of 100%, six subcategories (red) were unusable for analysis and model training.
Table 2. Response rate derived from tumor diameter. Response rate is presented by percentages and proportions. With a response rate of 100%, six subcategories (red) were unusable for analysis and model training.
.Week 4Week 8Week 12
Pleura83% (39/47)95% (36/38)97% (28/29)
Lymph nodes80% (12/15)91% (10/11)100% (10/10)
Peritoneum100% (12/12)100% (8/8)100% (2/2)
Soft tissues100% (9/9)100% (7/7)71% (5/7)
All87% (72/83)95% (61/64)94% (45/48)
Table 3. Response rate derived from tumor volume. Response rate is presented as percentages and proportions. Because of its limited sample size, one subcategory was unusable for analysis and model training (red).
Table 3. Response rate derived from tumor volume. Response rate is presented as percentages and proportions. Because of its limited sample size, one subcategory was unusable for analysis and model training (red).
.Week 4Week 8Week 12
Pleura30% (14/47)32% (12/38)28% (8/29)
Lymph nodes50% (7/15)55% (6/11)33% (3/10)
Peritoneum33% (4/12)25% (2/8)50% (½)
Soft tissues33% (3/9)28% (2/7)15% (1/7)
All34% (28/83)34% (22/64)27% (13/48)
Table 4. Response rate of Mean SUV tumors responses. Response rate is presented as percentages and proportions. Because of their limited sample size or inadequate response rate, some subcategories were unusable for analysis and model training (red).
Table 4. Response rate of Mean SUV tumors responses. Response rate is presented as percentages and proportions. Because of their limited sample size or inadequate response rate, some subcategories were unusable for analysis and model training (red).
.Week 4Week 8Week 12
Pleura20.5% (8/39)58% (7/12)36% (9/25)
Lymph nodes64% (7/11)100% (1/1)14% (1/7)
Peritoneum40% (4/10)34% (2/6)0% (0/1)
Soft tissues0% (0/7)0% (0/1)0% (0/7)
All67% (19/67)50% (10/20)25% (10/40)
Table 5. Radiomics associated with diameter-based response of target tumor. Number of radiomic features that had a significant difference of means between the responder/non-responder (Kruskal–Wallis test). Some subcategories were not evaluated (NA) because of the limited response rate.
Table 5. Radiomics associated with diameter-based response of target tumor. Number of radiomic features that had a significant difference of means between the responder/non-responder (Kruskal–Wallis test). Some subcategories were not evaluated (NA) because of the limited response rate.
OrganRadiomicsΔRadiomics
W4W8W12W8W12
Pleura15 (N = 47)7 (N = 38)0 (N = 29)81 (N = 38)0 (N = 29)
Lymph nodes76 (N = 15)0 (N = 11)NA0 (N = 11)NA
PeritoneumNANANANANA
Soft tissuesNANA1 (N = 7)NA0 (N = 7)
Table 6. Radiomic features associated with volume-based response of target tumors. Number of radiomic features that had a significant difference of means between the responder/non-responder (Kruskal–Wallis test). Some subcategories were not evaluated (NA) because of the limited response rate.
Table 6. Radiomic features associated with volume-based response of target tumors. Number of radiomic features that had a significant difference of means between the responder/non-responder (Kruskal–Wallis test). Some subcategories were not evaluated (NA) because of the limited response rate.
OrganRadiomicsΔRadiomics
W4W8W12W8W12
Pleura185 (N = 47)584 (N = 38)290 (N = 29)505 (N = 38)221 (N = 29)
Lymph nodes99 (N = 15)93 (N = 11)53 (N = 10)404 (N = 11)164 (N = 10)
Peritoneum122 (N = 12)55 (N = 8)NA0 (N = 8)0 (N = 2)
Soft tissues129 (N = 9)2 (N = 7)0, (N = 7)2 (N = 7)0 (N = 7)
Table 7. Radiomic features associated with PERCIST response of target tumors.
Table 7. Radiomic features associated with PERCIST response of target tumors.
OrganRadiomicsΔRadiomics
W4W12W4W12
Pleura532 (N = 39)209 (N = 25)73 (N = 39)71 (N = 25)
Lymph nodes282 (N = 11)53 (N = 7)48 (N = 11)NA
Peritoneum89 (N = 10)NA278 (N = 10)NA
Soft tissuesNANANANA
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Beaumont, H.; Iannessi, A.; Thinnes, A.; Jacques, S.; Quintás-Cardama, A. Radiomics-Based Prediction of Treatment Response to TRuC-T Cell Therapy in Patients with Mesothelioma: A Pilot Study. Cancers 2025, 17, 463. https://doi.org/10.3390/cancers17030463

AMA Style

Beaumont H, Iannessi A, Thinnes A, Jacques S, Quintás-Cardama A. Radiomics-Based Prediction of Treatment Response to TRuC-T Cell Therapy in Patients with Mesothelioma: A Pilot Study. Cancers. 2025; 17(3):463. https://doi.org/10.3390/cancers17030463

Chicago/Turabian Style

Beaumont, Hubert, Antoine Iannessi, Alexandre Thinnes, Sebastien Jacques, and Alfonso Quintás-Cardama. 2025. "Radiomics-Based Prediction of Treatment Response to TRuC-T Cell Therapy in Patients with Mesothelioma: A Pilot Study" Cancers 17, no. 3: 463. https://doi.org/10.3390/cancers17030463

APA Style

Beaumont, H., Iannessi, A., Thinnes, A., Jacques, S., & Quintás-Cardama, A. (2025). Radiomics-Based Prediction of Treatment Response to TRuC-T Cell Therapy in Patients with Mesothelioma: A Pilot Study. Cancers, 17(3), 463. https://doi.org/10.3390/cancers17030463

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop