Radiomics and Machine Learning Analysis Based on Magnetic Resonance Imaging in the Assessment of Colorectal Liver Metastases Growth Pattern

To assess Radiomics and Machine Learning Analysis in Liver Colon and Rectal Cancer Metastases (CRLM) Growth Pattern, we evaluated, retrospectively, a training set of 51 patients with 121 liver metastases and an external validation set of 30 patients with a single lesion. All patients were subjected to MRI studies in pre-surgical setting. For each segmented volume of interest (VOI), 851 radiomics features were extracted using PyRadiomics package. Nonparametric test, univariate, linear regression analysis and patter recognition approaches were performed. The best results to discriminate expansive versus infiltrative front of tumor growth with the highest accuracy and AUC at univariate analysis were obtained by the wavelet_LHH_glrlm_ShortRunLowGray Level Emphasis from portal phase of contrast study. With regard to linear regression model, this increased the performance obtained respect to the univariate analysis for each sequence except that for EOB-phase sequence. The best results were obtained by a linear regression model of 15 significant features extracted by the T2-W SPACE sequence. Furthermore, using pattern recognition approaches, the diagnostic performance to discriminate the expansive versus infiltrative front of tumor growth increased again and the best classifier was a weighted KNN trained with the 9 significant metrics extracted from the portal phase of contrast study, with an accuracy of 92% on training set and of 91% on validation set. In the present study, we have demonstrated as Radiomics and Machine Learning Analysis, based on EOB-MRI study, allow to identify several biomarkers that permit to recognise the different Growth Patterns in CRLM.


Introduction
Colorectal cancer (CRC) is one of the main common cancer worldwide, representing about the 10% of new detected cancers in 2020 [1]. Moreover, it is estimated that in 2040 its prevalence will rise quickly to >3 million cases per year [2,3]. Although the management of patients within a multidisciplinary team have improved the clinical outcome thanks to a closer patient follow-up to obtain an earlier detection of metastatic disease and an improvement in the efficacy of systemic therapies, and of surgical indications and -quality, based on a better patient selection, however metastatic spread is the main cause of death [4][5][6][7][8].
In addition, at the diagnosis, about 20% of patients have liver metastases (CRLM), while approximately 40-50% of patients will develop CRLM during surveillance, either after primary tumor resection/multimodal therapy or after the resection of CRLM [8][9][10][11][12][13][14][15][16]. Another key is related to the fact that about 60% of patients can expect to have liver recurrence even after a R0 resection of the primary CRLM. Several researches have demonstrated as patients with (a) T3/T4 CRC status, (b) local positive node, and more than 3 CRLM have higher possibility of recurrence [17][18][19]. In addition, several researches have suggested that the distinct growth pattern of CRLM could be linked to liver recurrence and overall survival (OS) after conversion therapy and resection [20][21][22]. With the increasing frequency of CRLM resections, the histology of the front morphology of CRLM is accessible and different growth patterns have been reported: in the "desmoplastic" or "encapsulated" model, tumour cells are separated from the liver parenchyma by a fibrotic stroma border that is distinguished by a 'pushing' or 'expansive' pattern, in which the hepatic plaques adjacent to the metastases are flattened with no intermediate fibrotic tissue. The model called "invasive" or "replacement" in which cancer cells infiltrate normal surrounding liver parenchyma, replace the hepatocytes and the sinusoidal stromal [20][21][22]. A recent review [22] assessed 17 clinical studies, including either chemonaïve and chemotherapytreated patients, showed that in the 82.4% a statistically significant favourable outcome was reported for patients with desmoplastic CRLM while in the 66.7% a significantly unfavourable outcome for patients with a predominantly replacement-type CRLM was demonstrated [22].

Materials and Methods
Local Ethical Committee board accepted this retrospective study, renouncing to the patient informed consent. Patients were selected by radiological database considering the temporal period from January 2018 to June 2021 and the following inclusion criteria: (1) liver pathological proven metastases; (2) MRI study of high quality in pre-surgical setting and (3) a follow-up Computed Tomography (CT) scan of at least six months after surgical liver resection.
The patient cohort included a training set of 51 patients with 121 liver metastases, and an external validation set of 30 patients with a single lesion. The validation cohort was provided by "Careggi Hospital", Florence, Italy. The patient characteristics were reported in Table 1.

MR Imaging Protocol
MR studies were performed with two 1.5T MR scanners: Magnetom Symphony (Siemens, Erlangen, Germany) and Magnetom Aera (Siemens).
The MRI study protocol included conventional sequences, T1 weighted (W), without contrast medium administration, and T2-W, Diffusion Weighted Imaging (DWI) with seven b values in order to obtain functional parameters with mono-exponential model and T1-W sequences after the administration of contrast medium. In Table 2 we reported MR study protocol. According to the different phase of patient management, our study protocol includes the possibility to administrate a liver-specific contrast agent (in pre surgical setting) and a non-liver-specific contrast agent (in characterization and staging phase). In this study we assessed images obtained employing a liver-specific contrast agent (0.1 mL/kg of Gd-EOB-BPTA-Primovist, Bayer Schering Pharma, Berlin, Germany). A power injector (Spectris Solaris ® EP MR, MEDRAD Inc., Indianola, IA, USA) was used to administrate the contrast agent using an infusion rate of 2 mL/s.

MRI Post-Processing
For each volume of interest, 851 radiomic features were extracted as median values on the segmented volume by two expert radiologists in abdominal imaging and MRI, using the PyRadiomics tool [61] and as previously reported in [62,63] and as reported in [https: //readthedocs.org/projects/pyradiomics/downloads/ (accessed on 21 December 2021)].

Reference Standard
Histopathologic data, from routine report were used as the reference standard for determining metastases growth pattern. Lesions with desmoplastic growth pattern were defined as expansive; lesions with predominantly replacement-type were defined infiltrative.

Statistical Analysis
Intraclass correlation coefficient was used to evaluate the variability by two radiologists. The non-parametric Kruskal-Wallis test was performed to recognize differences statistically significant of radiomics metrics median values in the identification of tumor growth front (expansive versus infiltrative).
Chi square test was used to evaluate if the tumor growth front (expansive versus infiltrative) was correlated with the local recurrence occurrence.
Receiver operating characteristic (ROC) and Youden index was used to calculate the optimal cut-off for each metric and area under curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV) and accuracy.
A feature selection was made to delete the redundant and non informative metrics considering the following findings: metrics significant by Kruskal-Wallis test and with an accuracy when considered alone major of 75%. A linear regression analysis was made to calculate the linear regression model of all significant metrics.
Moreover, approaches of artificial intelligence were considered to define the best classifier of all significant metrics. The classifiers considered included support vector machine (SVM), k-nearest neighbors (KNN), artificial neural network (NNET), and decision tree (DT)) [63]. The best model was selected considering the highest area under ROC curve and the highest accuracy. Each classifier was trained with a 10-k fold cross validation and was validated using the external validation set.
McNemar test was used to evaluate that the results of the dichotomy tables were statistically significant.
A p-value < 0.05 was considered as significant. The analysis was made considering the Statistics and Machine Toolbox of MATLAB R2021b (MathWorks, Natick, MA, USA).

Results
No correlation statistically significant was relived between the tumor growth front and the local recurrence occurrence (p-value = 0.82 at Chi square test).
On univariate analysis (Table 3), a variable number of metrics were statistically significant which were different when extracted from different MR sequences: 15 significant predictors extracted from T2W SPACE; 8 significant predictors extracted from the arterial phase; 9 significant predictors extracted from the portal phase; 8 significant predictors extracted from the EOB phase.
The best results (Table 3) with the highest accuracy and AUC at univariate analysis to discriminate expansive versus infiltrative front of tumor growth were obtained by the wavelet_LHH_glrlm_ShortRunLowGrayLevelEmphasis from portal phase sequence with accuracy of 82%, sensitivity of 84%, specificity of 77%, PPV and NPV of 85% and 74%, respectively, and a cut-off value of 0.12. The results were statistically different from those obtained with the T2W and arterial phase MR sequences while were to those obtained from EOB phase sequence with the wavelet_HLL_glcm_InverseVariance that obtained and AUC of 78% and an accuracy of 83%. Linear regression model increased the performance obtained respect to the univariate analysis (see Table 4) for each sequence except that for EOB-phase sequence. The best results were obtained by a linear regression model of 15 significant features extracted by the T2W SPACE sequence with accuracy of 90%, a sensitivity of 95%, a specificity of 95%, a PPV and a NPV of 80% and 89% respectively. These results were statistically different from the univariate analysis results and compared with the results of metrics extracted from other MR sequences (p-value < 0.01 on McNemar's test). Table 5 reported the coefficients of metrics and intercept of the best linear regression model. Furthermore, using pattern recognition approaches, the diagnostic performance to discriminate the expansive versus infiltrative front of tumor growth increased again and the best classifier was a weighted KNN trained with the 9 significant metrics extracted from the portal phase sequence obtaining 92% accuracy on training set and 91% at validation set with an AUC of 0.97 and 0.99, respectively analysis (see Table 6). These results were statistically different from the univariate analysis results and linear regression analysis (p-value < 0.01 on McNemar's test).

Discussion
Since liver is the most common site of distant metastases in patients with CRC and, surgical resection is the only curative treatment, it is clear that the possibility to identify several prognostic features of CRLM to guide the proper patient management, remains a critical key open question. In addition, according to current principles for oncologic liver All results of the dichotomy tables were statistically significant (p-value < 0.01 at McNemar test).

Discussion
Since liver is the most common site of distant metastases in patients with CRC and, surgical resection is the only curative treatment, it is clear that the possibility to identify several prognostic features of CRLM to guide the proper patient management, remains a critical key open question. In addition, according to current principles for oncologic liver surgery, liver resection aims to remove the lesion with adequate margins and with sufficient liver remnant volume. This needs a multi-parametric patient assessment for a correct evaluation of the lesion and functional liver status [56][57][58]. Several features could guide the surgical procedure, comprising liver functional status, patient general status, tumour size, localization, and vascular infiltration.
Also, today, several researches have suggested that the distinct CRLM growth pattern is associated with differences in tumor local recurrence and OS. The rationale behind the need to know CRLM growth patterns is at least twofold: on one hand, the patterns may be useful for prognostication and may aid treatment decisions; on the other hand, it is likely that understanding these distinct patterns of metastatic progression will provide important insights into the biological mechanisms that support tumour growth in the liver [22]. Fernández Moro et al. [22] reported that microvessel co-option, besides providing vascular supply, renders metastases of the replacement type resistant to anti-angiogenic therapy. Furthermore, the authors suggested an alternative approach for treating replacement-type metastases. It is worth noting that several studies published before FDA approval of anti-angiogenic therapy found a favourable prognosis for patients with desmoplastic-type CRLM, suggesting that there are important mechanistic insights beyond differences in angiogenesis to be gained from the different growth patterns [22].
In this scenario, it is clear that having the possibility, during an imaging study, to identify biomarkers that could correlate with the growth of the lesion, allows for better treatment selection [58][59][60]. At the best of our knowledge no one study has evaluated the Radiomics and Machine Learning Analysis Based on MRI study, in the assessment of CRLM growth pattern. With regard to univariate analysis, a variable number of metrics were statistically significant: 15 extracted from T2W SPACE; 8 from the arterial phase; 9 extracted from the portal phase; 8 extracted from the EOB phase. The best results with the highest accuracy and AUC at univariate analysis to discriminate expansive versus infiltrative front of tumor growth were obtained by the wavelet_LHH_glrlm_ShortRunLowGrayLevelEmphasis from portal phase.
With regard to linear regression model, this increased the performance obtained respect to the univariate analysis for each sequence except that for EOB-phase sequence. The best results were obtained by a linear regression model of 15 significant features extracted by the T2W SPACE. These results were statistically different from the univariate analysis results. Furthermore, using pattern recognition approaches, the diagnostic performance to discriminate the expansive versus infiltrative front of tumor growth increased again and the best classifier was a weighted KNN trained with the 9 significant metrics extracted from the portal phase. Our results were confirmed by validation external dataset.
However, when we evaluated the tumor growth pattern and the local recurrence, we found no statistically significant correlation. This result could be explained by the number of patients under examination and the short follow-up period.
This study has several weaknesses: (1) the small sample size, although the analysis was done on a homogeneous group and on all single lesion; (2) the retrospective nature, (3) a manual segmentation. Furthermore, we not evaluated: (4) the impact of chemotherapy on our data, while we assessed all single study protocol phase, in order to identify the more accurate sequence in the assessment of growth pattern and our results were confirmed by a validation external dataset.

Conclusions
In the present study, we have demonstrated as Radiomics and Machine Learning Analysis, based on EOB-MRI study, allow to identify several biomarkers that permited to recognise the different Growth Patterns in CRLM. The best results with the highest accuracy and AUC at univariate analysis were obtained by the wavelet_LHH_glrlm_ShortRunLowGrayLevelEmphasis from portal phase. Linear regression model increased the performance obtained respect to the univariate analysis for each sequence, except that for EOB-phase sequence. Furthermore, using pattern recognition approaches, the best classifier was a weighted KNN trained with the 9 significant metrics extracted from the portal phase.  Informed Consent Statement: This study aligned with National appropriate guidelines and procedures. Renouncing the need for informed patient consent given the study nature.