Next Article in Journal
Enhancing Diagnostic Accuracy of Neurological Disorders Through Feature-Driven Multi-Class Classification with Machine Learning
Previous Article in Journal
Genomic Insights into Tumorigenesis in Newly Diagnosed Multiple Myeloma
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Diagnostics
Volume 15
Issue 17
10.3390/diagnostics15172131
diagnostics-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

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

Transformer-Based HER2 Scoring in Breast Cancer: Comparative Performance of a Foundation and a Lightweight Model

by
Yeh-Han Wang
1,2,
Min-Hsiang Chang
3,*,
Hsin-Hsiu Tsai
4,
Chun-Jui Chien
4 and
Jian-Chiao Wang
4
1
Department of Anatomical Pathology, Fu Jen Catholic University Hospital, Fu Jen Catholic University, New Taipei City 24352, Taiwan
2
School of Medicine, College of Medicine, Fu Jen Catholic University, New Taipei City 24205, Taiwan
3
Li Jen Pathology Clinic, Taipei 11469, Taiwan
4
AI Lab., Quanta Computer Inc., Taoyuan City 33377, Taiwan
*
Author to whom correspondence should be addressed.
Diagnostics 2025, 15(17), 2131; https://doi.org/10.3390/diagnostics15172131 (registering DOI)
Submission received: 5 July 2025 / Revised: 21 August 2025 / Accepted: 22 August 2025 / Published: 23 August 2025
(This article belongs to the Section Machine Learning and Artificial Intelligence in Diagnostics)
Browse Figures
Review Reports Versions Notes

Abstract

Background/Objectives: Human epidermal growth factor 2 (HER2) scoring is critical for modern breast cancer therapies, especially with emerging indications of antibody–drug conjugates for HER2-low tumors. However, inter-observer agreement remains limited in borderline cases. Automatic artificial intelligence-based scoring has the potential to improve diagnostic consistency and scalability. This study aimed to develop two transformer-based models for HER2 scoring of breast cancer whole-slide images (WSIs) and compare their performance. Methods: We adapted a large-scale foundation model (Virchow) and a lightweight model (TinyViT). Both were trained using patch-level annotations and integrated into a WSI scoring pipeline. Performance was evaluated on a clinical test set (n = 66), including clinical decision tasks and inference efficiency. Results: Both models achieved substantial agreement with pathologist reports (linear weighted kappa: 0.860 for Virchow, 0.825 for TinyViT). Virchow showed slightly higher WSI-level accuracy than TinyViT, whereas TinyViT reduced inference times by 60%. In three binary clinical tasks, both models demonstrated a diagnostic performance comparable to pathologists, particularly in identifying HER2-low tumors for antibody–drug conjugate (ADC) therapy. A continuous scoring framework demonstrated a strong correlation between the two models (Pearson’s r = 0.995) and aligned with human assessments. Conclusions: Both transformer-based artificial intelligence models achieved human-level accuracy for automated HER2 scoring with interpretable outputs. While the foundation model offers marginally higher accuracy, the lightweight model provides practical advantages for clinical deployment. In addition, continuous scoring may provide a more granular HER2 quantification, especially in borderline cases. This could support a new interpretive paradigm for HER2 assessment aligned with the evolving indications of ADC.

Graphical Abstract

1. Introduction

Breast cancer is one of the most common malignancies worldwide and remains a leading cause of cancer-related death among women. Human epidermal growth factor 2 (HER2) immunohistochemistry (IHC) scoring plays a pivotal role in guiding targeted therapy decisions. According to the ASCO/CAP guidelines, patients with HER2-positive tumors (3+) are eligible for anti-HER2 therapies, such as trastuzmab, whereas equivocal cases (2+) require confirmatory in situ hybridization (ISH) testing to assess HER2 gene amplification. In recent years, the development of novel anti-HER2 agents, particularly antibody–drug conjugates (ADCs), has expanded treatment options. Clinical trials have demonstrated that ADCs significantly prolong progression-free and overall survival in patients with HER2-low tumors (1+ or IHC 2+ without amplification), which account for approximately 45–64% of all breast cancers [1,2,3]. Moreover, ADCs are now being explored in tumors with ultralow HER2 expression (defined as 0 < HER2 < 1+) [4,5]. However, the expanding therapeutic landscape has also intensified a longstanding challenge: the inherent subjectivity and limited consistency of HER2 scoring among pathologists, especially in borderline categories.
HER2 scoring is influenced by multiple factors, including fixation time, antibody clones, staining protocols, tumor heterogeneity, and most importantly, inter-observer variability [2,6,7]. When using standardized specimens, identical tissue sections, or whole-slide images (WSIs), inter-observer agreement for HER2 IHC scoring can reach 86–92%, with Cohen’s kappa values between 0.7 and 0.8 [8]. However, most validation studies and international proficiency assessments were primarily designed to evaluate HER2-positive (3+) detection. Recent investigations focusing on HER2-low diagnoses have revealed greater variability in distinguishing HER2-low from HER2-negative, with agreement rates ranging from 33% to 80% across studies [4,8,9,10,11]. Due to these discrepancies, reader retraining was often required to achieve diagnostic alignment. Furthermore, in the context of HER2-ultralow detection, one study reported low binary consistency between 0 and ultralow classifications, with a Fleiss kappa of just 0.292 [12]. These variations highlight the urgent need for more reproducible scoring methods, as inconsistent classification results may propagate into both treatment decisions and the design of ADC-based studies. In this context, artificial intelligence (AI)-based approaches may offer complementary support for improving the consistency and efficiency of HER2 scoring.
Although the first FDA-cleared image analysis algorithm for HER2 scoring was launched in 2009, automatic HER2 assessment remains one of the most actively explored topics in digital pathology. Over the past decade, advances in deep learning have markedly improved the precision and robustness of image analysis models. These approaches have employed various inference levels—including cell-, area-, patch-, and WSI-based designs—and typically generate quantitative results, along with heatmaps to aid visual verification [6,13,14,15,16,17,18]. Since the introduction of the HER2-low classification, several studies have demonstrated that AI-assisted scoring can enhance intra-observer consistency, particularly in borderline cases [7,19,20,21,22]. By quantifying the proportion of cells with different staining intensities, AI tools can help pathologists achieve more reproducible evaluations and may also support junior readers in aligning with expert-level interpretations. However, most prior approaches have adopted convolutional neural networks as the primary architecture, whereas more recent transformer-based models are designed to incorporate adjacent contextual information, mirroring how pathologists integrate surrounding features into diagnostic decisions. Recent advances in lightweight transformers (TinyVit) have demonstrated competitive performance in general image classification benchmarks, suggesting their potential for efficient deployment in specialized domains [23].
Pathology foundation models utilize large-scale, transformer-based AI architectures that are pretrained on vast WSI datasets across multiple organ systems [24,25,26,27]. The first published example, CTransPath, was introduced in 2022 and was trained on publicly available datasets, such as The Cancer Genome Atlas (TCGA) and the Pathology AI Platform (PAIP). In 2023, Virchow was released, a vision transformer-based model with 632 million parameters, pretrained on approximately 1.5 million WSIs from 119,629 patients at Memorial Sloan Kettering Cancer Center [26]. By leveraging exposure to diverse morphologic patterns, foundation models have demonstrated stronger generalization capabilities and superior performance in zero-shot or few-shot tasks than conventional neural networks. Also, these foundation models have outperformed traditional conventional neural network approaches across multiple tasks [24,25,26,27]. These studies indicated that transformer-based architectures can achieve better generalization and contextual understanding than convolutional approaches in histopathology applications. However, these performance gains come at the cost of substantial computational requirements. The size and complexity of foundation models pose practical limitations for clinical deployment, where inference speed, hardware efficiency, and scalability are critical considerations. As a result, compact transformer models with fewer parameters, which require less computing power, may provide a more feasible alternative for routine diagnostic workflows.
Therefore, we adapted two transformer-based models for automated HER2 scoring in breast-cancer WSIs. One is a state-of-the-art large-scale foundation model (Virchow), representing the upper bound of current AI performance with extensive pretraining. The other is a lightweight transformer (TinyViT) with 21 million parameters, offering a computationally efficient alternative that may be more practical for on-premises deployment in clinical pathology settings. Their performance was evaluated across multiple levels, including patch-level classification, WSI-level scoring, and inference efficiency. In addition, we explored a continuous scoring approach to complement conventional HER2 categorization and assess its potential interpretive value.

2. Material and Methods

2.1. Case Collection

In this study, 108 breast cancer HER2 slides were included. Paraffin blocks were retrieved from the Li Jen Pathology Clinic between June 2022 and December 2023 (Table 1). Of the included slides, 25 (23%) were HER2 0, 36 (33%) were HER2 1+, 29 (27%) were HER2 2+, and 18 (17%) were HER 3+ according to the original pathology reports. Forty-nine slides were used to build the model at the patch level, and 66 slides (including 7 test slides also used for the patch classifier task) were used as a test set to evaluate WSI-level scoring. This study received approval from the Research Ethics Review Committee of the Far Eastern Memorial Hospital (No. 113141-E, date of approval: 5 June 2024). All blocks used in the clinical diagnosis setting were retrieved from the repository. These slides and images were anonymized and did not contain any personal information. Therefore, the requirement for informed consent was waived.

2.2. Specimen Staining and Image Acquisition

The retrieved paraffin-embedded tissue blocks had been fixed in 10% neutral formalin for 6–72 h before further processing. The blocks were cut into 5 μm thick sections and mounted on hydrophilic slides. The slides were stained with ready-to-use anti-HER2 antibodies (4B5, Ventana Medical Systems, Oro Valley, AZ, USA) using a standardized Roche protocol on a Benchmark GX (Ventana Medical Systems), ensuring consistent staining conditions across all specimens. Then, the slides were scanned with a Hamamatsu S210 (Hamamatsu, Iwata City, Shizuoka, Japan) at 40× magnification (0.23 μm per pixel resolution), and these images were saved in NDPI format.

2.3. Data Preparation and Model Training

To quantify HER2’s intensity and visualize distributions as heatmaps, the AI workflow was based on an image patch classifier (Figure 1) [6]. First, a pathologist (Dr. A) annotated invasive carcinomas in WSIs, and the tumor areas were partitioned into image patches (256 × 256 pixels, equivalent to 60 μm × 60 μm, containing 10 to 30 cells). Then, Dr. A labeled 5109 patches (Table 2) into five categories: non-tumor, 0 (no staining), 1+ (more than 10% of cells with weak staining), 2+ (more than 10% of cells with moderate incomplete membranous staining), and 3+ (more than 10% of cells with strong complete membranous staining). During patch-level annotation, any patches containing staining artifacts or technical irregularities were systematically labeled as ‘non-tumor’ to strengthen the model’s robustness against similar technical variations. After a washout period of 8 weeks, Dr. A blindly relabeled 1966 patches without any previous labeling information to evaluate intra-observer variation, and the other pathologist (Dr. B) labeled a portion of the data (1200 patches) to evaluate inter-observer variation. This assessment was used to determine a reasonable endpoint for the algorithm training, because it was implausible for models to outperform intra-observer consistency.
Two transformer-based neural networks were adopted in this study (Table S1). The first was Virchow, a state-of-the-art pathology foundation model with 632 million parameters based on the Vision Transformer architecture [26]. The second was TinyViT, a lightweight vision transformer model comprising 21 million parameters [27]. Both networks were trained on a dataset of image patches labeled by Dr. A to build HER2 patch classifiers, referred to as Model V (Virchow) and Model T (TinyViT). No architectural modifications were made, nor were layers frozen to the base models.
After reaching the designated training endpoints, the models were integrated into a WSI HER2 scoring workflow. Tumor regions within each HER2-stained WSI were first segmented by a model, which had been fine-tuned from our previously developed virtual cytokeratin algorithm [28]. This cytokeratin-supervised model predicted breast epithelial components, especially tumor parts, in IHC stains. By constraining patch classifiers to tumor areas, scoring accuracy improved, while computational load decreased, and artifact robustness was enhanced by excluding most non-tissue processing artifacts. Upon completion of inference, the system outputs the proportion of patches in each HER2 intensity level (P0 to P3+), the final WSI-level HER2 score, and a visual heatmap overlay for interpretability. According to the ASCO/CAP guidelines, HER2 scoring is based on the percentage of tumor cells with specific membranous staining intensities. In our workflow, the proportion of patches classified into each intensity category was used as a proxy for the percentage of stained tumor cells, following the same approach that leverages patch-level predictions for WSI-level scoring. Specifically, our approach is analogous to that of Pham et al. [6], who employed a weakly supervised deep learning framework to create an interpretable HER2 scoring model by aligning its outputs with ASCO/CAP clinical guidelines. In this phase, 66 WSIs were used to evaluate accuracy and consistency with the ground truth from pathology reports. Since the workflow quantified the area of each staining intensity, we also calculated a continuous HER2 score (c-score = 0 × P0 + 1 × P1+ + 2 × P2+ + 3 × P3+) by weighted average of patch-level classification. Regarding computing power demand by both models, a computer with RTX4060 Ti GPU (16GB) was used to compare the inference times for five WSIs between the models.

2.4. Statistical Analysis

Python software (ver. 3.8.13; https://www.python.org/) and SPSS version 21 for Windows (Chicago, Armonk, NY, USA) were used to compute the performance metrics of the two models. The metrics included accuracy, recall, precision, and F1 score in patch-level classifications. Consistency between models and pathologists was evaluated by calculating the linear weighted kappa (LWK). Performance in WSI HER2 scoring was also assessed by accuracy and LWK. The c-score correlation between the two models was calculated using Pearson’s correlation coefficient (r), and the inference time comparison was analyzed using the paired t-test. Statistical significance was set at p < 0.05.

3. Results

3.1. Intra- and Inter-Observer Consistency at the Patch Level

Dr. A relabeled approximately 40% of the dataset (n = 1966) after a washout period of 8 weeks. The results showed an accuracy of 76.2% and good consistency (LWK = 0.825) between the two sets of intra-observer labels (Figure 2A). The most disparate classes were non-tumor and HER2 1+. If non-tumor patches (n = 1200) were excluded, the accuracy increased to 85.6%, and the LWK value rose to 0.879. Inter-observer agreement between the two pathologists (Dr. A vs. Dr. B) showed an overall accuracy of 58.5% and an LWK value of 0.611. After excluding non-tumor regions, accuracy improved to 61.9% and LWK to 0.671, consistent with prior reports of variability in HER2 IHC interpretation [6]. Therefore, we determined an accuracy of 70% or an LWK of 0.7 as our training endpoint.

3.2. Performance of the Two Models at the Patch Level

To evaluate the training stability and optimization behavior of the two backbone models, we compared their validation F1 scores and accuracy curves throughout training (Figure S1). Model V achieved a higher performance in earlier epochs, with a peak F1 score (0.771) at epoch 13 and the highest validation accuracy (79.4%) at epoch 5, followed by mild fluctuations throughout the training process. In contrast, Model T demonstrated a slow, gradual, but stable improvement over the training epochs, eventually reaching its peak F1 score (0.755) and validation accuracy (76.1%) at epoch 40.
In the test set, Model V reached an overall accuracy of 73.3% and an LWK of 0.738 (Figure 2C). Accuracy and LWK increased to 79.6% and 0.844 by excluding non-tumor patches, as Model V had difficulties in non-tumor patches and misrecognized some patches as 0 or 1+. Moreover, it tended to upgrade 1+ to 2+ and 2+ to 3+. Model V’s misclassification patterns can be attributed to its higher sensitivity to color intensity variations. This large foundation model appears to prioritize staining intensity over morphological context, leading to misidentification of non-tumor regions with background staining as weak positive (0 or 1+) patches and systematic upgrading of patches based primarily on color intensity. Although Model T had fewer parameters, it outperformed Model V and reached an overall accuracy of 76.9% and an LWK of 0.787 (Figure 2D). Without non-tumor patches, accuracy and LWK improved to 81.4% and 0.857, respectively. Model T also tended to upgrade patches, but it performed more effectively in non-tumor regions. Like for human raters, 1+ and 2+ patches were most difficult to assess for both models.
Overfitting assessment based on training dynamics (Figure S1) showed stable validation performance without significant degradation relative to training accuracy. Both models demonstrated appropriate convergence patterns: Model V reached a plateau early with mild fluctuations, while Model T showed gradual, stable improvement. Validation curves indicated generalization without excessive training data memorization. Furthermore, the final patch-level performance’s positioning between inter-observer (LWK = 0.611) and intra-observer agreement (LWK = 0.825) suggests realistic learning within human performance boundaries.

3.3. Performance of the Two Models at the WSI Level

In WSI HER2 scoring (Figure 3A,B), Model V achieved a higher accuracy (83.3%) than Model T (78.8%). Regarding agreement with ground truth, Model V also performed better (LWK = 0.860) than Model T (LWK = 0.825). However, the difference in performance was not significant (p = 0.321), and the assessments by both models were equivalent to those by pathologists [9,22]. Model V and Model T misclassified 11 and 14 cases, respectively, sharing errors on eight slides (5 cases were 2+, 2 were 1+, and 1 was 0). All misclassified cases were off by one grade from the ground truth. Model T exhibited a balanced distribution of over- and underestimation, but Model V tended to underestimate (score 2+ to 1+), because more non-tumor patches were classified as 0 or 1+ (Figure 3C).
To assess the performance for clinical decision-making, three clinically relevant binary classification tasks were evaluated (Table 3): (1) identifying HER2 IHC 3+ cases eligible for anti-HER2 therapy, (2) stratifying 2+ cases requiring confirmatory ISH, and (3) differentiating HER2-low (1+/2+) from HER2-negative (0) cases. In the first binary setting (3+ vs. 0/1+/2+), Model V showed perfect accuracy (100%), whereas Model T achieved an accuracy of 95.5% with a sensitivity of 100%. After excluding 3+ cases, the second binary task evaluated the ability to distinguish 2+ from 0/1+. Model V reached an accuracy of 84.6% and a sensitivity of 55.6%, with eight missed cases. Model T achieved an accuracy of 80.8% and the same sensitivity of 55.6%, with two false positives and eight false negatives. In the last task for HER2-low identification (1+/2+ vs. 0), Model V, with an accuracy of 94.2% and a sensitivity of 94.6%, outperformed Model T, which only achieved an accuracy of 86.5% and a sensitivity of 86.5%. Both models were comparable to pathologists and state-of-the-art algorithms in other studies [6,8,9,12,20].
In terms of continuous HER2 score (c-score), the two models strongly correlated (Pearson’s r = 0.995) across all 66 test cases (Figure 4). This suggests that they were highly consistent in interpreting the overall HER2 staining intensity at the WSI level, although their tiered HER2 scores might differ. Furthermore, their c-scores were visually associated with ground-truth HER2 scores.

3.4. Inference Time for the Two Models

Runtime analysis was conducted on five test cases to compare the inference efficiency between the two models. The average processing time for Model V was 1024.6 (688.7–1875.8) s per WSI. In contrast, Model T required on average only 368.7 (292.4–438.4) s, resulting in a time reduction of 60.7% (52.0–77.2%, p = 0.013). These results demonstrated that Model T saved substantial computing power and inference time.

4. Discussion

In this study, we trained two deep learning structures, a large foundation model (Virchow, Model V) and a lightweight transformer (TinyViT, Model T), for automatic HER2 scoring in breast cancer WSIs. Both models demonstrated high accuracy and substantial agreement with the pathologist-annotated ground truth in patch-level classification, WSI-level scoring, and binary decision categorization. While Model V showed slightly higher performance in WSI-level accuracy and agreement metrics, Model T achieved equivalent abilities with a significantly faster inference time. In addition, we introduced a continuous score framework for the two models to assess HER2 expression, and their results were highly consistent. These findings highlight the practical feasibility of both models in HER2 assessment and set the stage for deeper comparisons in learning dynamics, predictive behavior, and clinical utility.
This work was inspired by a previous study conducted by Pham et al., who demonstrated that patch-based HER2 evaluation can achieve high concordance with cell-based ground truth in digital pathology workflows [6]. We adopted patches instead of cells as a proxy to evaluate staining intensity and completeness. This eased the manual annotation task and allowed us to evaluate intra-observer consistency, inter-observer consistency, and model performance at a fundamental level. Consistency among pathologists is limited, especially in interpreting weak-to-moderate staining (1+ and 2+). Even for the same pathologists, overall accuracy was 76.2%, and accuracy for 1+ patches was only 62.7% in the relabeling task, which provided benchmarks to assess the models’ performance. During training, the large-scale foundation model, Virchow, learned quickly and reached a plateau in early epochs, corresponding to the characteristics of foundation models. By contrast, the lightweight Model T grew gradually but eventually achieved a performance equivalent to that of Model V. This pattern is consistent with the expectation that foundation models may generalize quickly in early stages, but lightweight models can also effectively adapt provided sufficient supervised training.
In patch-level classifications, both models exhibited intra-observer-level performance. The lightweight Model T achieved slightly higher overall accuracy than Model V, indicating better alignment with the manually labeled ground truth. However, like for humans, 1+ and 2+ patches were the most difficult classification items for both models, and both models tended to upgrade patches. After reviewing these misclassified patches, we identified several plausible causes (see Supplementary Figure S2 for examples). First, many cases were borderline or equivocal even to pathologists, particularly when distinguishing between 1+ and 2+ cases. Second, strong cytoplasmic staining, often seen in 1+ patches, may have misled models to overestimate their intensity. Third, some 2+ patches with strong staining but insufficient membranous completeness were incorrectly upgraded to 3+. These patterns indicate that both models were more sensitive to staining intensity than completeness, mirroring common interpretive challenges in manual HER2 assessment.
In WSI-level HER2 scoring, both models achieved performances similar to those of experienced pathologists and other state-of-the-art algorithms [6,19,20,29]. Model V demonstrated a marginally higher accuracy and LWK than Model T, although the difference was not significant. While both models exhibited a tendency to overestimate HER2 intensity at the patch level, their predictions diverged when scores were aggregated at the WSI level: Model T demonstrated a more balanced distribution of over- and underestimation, whereas Model V tended to predict lower scores, particularly in 2+ cases. This discrepancy may be attributed to differences in the processing of non-tumor regions. Although a tumor masking algorithm was applied to restrict inference to tumor areas, benign glands and stromal components were sometimes inevitably included. Both classifiers occasionally misclassified these regions as 0 or 1+, thereby reducing the relative proportion of 2+ patches (Figure 3C). This effect was more pronounced in Model V, which showed lower accuracy (66.3%) in identifying non-tumor patches. As a result, Model V produced more conservative HER2 scores than Model T at the WSI level, with clinical significance for borderline 2+ cases where small changes in patch proportion calculations can shift final classification below diagnostic thresholds.
To investigate clinical applicability, we further evaluated the models across three binary classification tasks that reflected common decision points in HER assessment. The first task focused on identifying HER 3+ cases eligible for anti-HER2 therapy. Both models performed excellently, and Model V even achieved perfect accuracy. These results suggest that both models are highly reliable in recognizing HER2 overexpression. The second task aimed to stratify 2+ cases that require confirmatory ISH testing. Both models reached moderate accuracy (~80–85%), but sensitivity was limited (55.6%), likely due to underestimation caused by misclassification of non-tumor regions as 0 or 1+. Several strategies could improve sensitivity for the identification of 2+ cases. For example, ensemble approaches combining both models could leverage their complementary strengths, as Model T demonstrated better non-tumor discrimination, while Model V showed superior overall accuracy. Additionally, confidence scoring mechanisms could flag borderline cases for manual review, particularly when patch-level predictions show high variance within WSIs. For such borderline cases, the manual review of HER2 heatmaps and verification of 2+ patch proportions were still necessary, highlighting the role of human-in-the-loop decision-making. The third task differentiated HER2-low from HER2-negative cases to identify potential candidates for ADC therapy. Both models performed similarly to experienced pathologists and previously published AI approaches [4,8,9,10,12]. Model V achieved a higher accuracy and sensitivity (~94%), whereas Model T reached ~86% in both parameters. These results indicate the feasibility of automated HER2 scoring as a decision-support tool, providing consistent, interpretable, and visually guided outputs for clinical integration. At the same time, these findings also reflect the inherent limitations of tiered scoring systems, particularly in borderline or heterogeneous cases [30,31].
The persistent challenges in 1+ and 2+ scoring reflect fundamental limitations at multiple analytical levels. At the patch level, 256 × 256 pixel patches face inherent difficulties in capturing membranous completeness and cellular proportion criteria that distinguish these borderline categories. Unlike 0 and 3+ cases with clear intensity differences, 1+ and 2+ assessments depend heavily on subtle morphological features that may be obscured at the current patch resolution. At the WSI level, borderline cases frequently contain heterogeneous patch populations with mixed staining patterns. When a WSI contains, for example, around 10% of 1+ or 2+ patches, the tier-based classification system forces a binary decision that may not reflect the underlying biological complexity. This aggregation challenge highlights the inherent limitations of discrete scoring systems for capturing the continuum of HER2 expression. To better capture the continuum of HER2 expression, we further explored a continuous scoring approach.
To provide a more nuanced quantification of HER2 expression, we introduced a continuous scoring framework (c-score) based on the proportion and intensity of predicted patch classes [32]. Despite architectural differences, the c-scores generated by Model V and Model T were highly correlated, suggesting consistent interpretation of HER2 expression at the WSI level. While their final categorical scores occasionally differed, the underlying c-scores reflected similar signal distributions, capturing the continuum of HER2 expression more faithfully than discrete labels. Moreover, c-scores exhibited a clear visual association with pathologist-assigned HER2 categories, offering an interpretable and intuitive measure that may aid in borderline or equivocal cases. This approach complements other strategies, such as the seven-class subdivision proposed by Pedraza et al. [33], which also aims to improve the handling of ambiguous cases by expanding the decision boundaries for the equivocal 2+ category before aggregating back to standard scores. Although further validation is needed, these findings highlight the potential of this c-score as a proof of concept for continuous HER2 assessment, suggesting a future direction for AI-assisted precision pathology.
In addition to accuracy and interpretability, inference efficiency is a practical consideration for clinical deployment of AI models. Runtime analysis revealed a substantial difference between the two models: the lightweight Model T completed WSI-level inference in approximately one-third of the time required by Model V. This significant reduction in processing time decreased computational resource demands and enhanced scalability in high-throughput clinical workflows. While foundation models like Model V can offer strong generalization and rapid early-stage convergence, their computational demands may pose challenges for routine use, particularly in resource-constrained settings. In contrast, lightweight models, such as Model T, demonstrate not only a competitive diagnostic performance but also superior inference efficiency, potentially lowering the barrier for real-world implementation in pathology laboratories. In this study, both models can operate on the same RTX4060 Ti GPU configuration (approximate workstation cost: USD 5000–8000), making the hardware barrier relatively low for clinical deployment. However, for high-throughput laboratories requiring faster Model V processing, upgrading to enterprise-grade GPUs (e.g., A100) may be considered, with infrastructure costs reaching USD 15,000–20,000. For clinical integration, our models are designed for compatibility with existing digital pathology workflows. The most effective implementation would involve integration with hospital information systems or established digital pathology platforms, enabling seamless incorporation into routine diagnostic workflows. Additionally, considering open-source model deployment could facilitate broader adoption across different institutional platforms and promote collaborative development. The interpretable outputs, including quantitative patch percentages and visual heatmaps, are designed to complement rather than replace pathologist expertise, supporting evidence-based diagnostic decisions within existing clinical frameworks.
Another promising future direction is to explore the feasibility of predicting HER2 status directly from H&E-stained images, a method proposed by researchers like Wang et al. [34]. Their model, named HAHNet, was designed to classify HER2 status using H&E-stained histological images, thereby potentially reducing the additional costs associated with IHC staining and alleviating the burden on pathologists. The study reported that their HAHNet model achieved superior performance compared to other computational methods for H&E-based prediction. While our study and many others focus on IHC-stained images, this alternative approach represents a valuable avenue for streamlining the diagnostic workflow and is a significant area for future research.
This study has some limitations. First, although our test set was derived from real-world diagnostic cases, the total number of cases (n = 66) was relatively small, and 7 cases for patch classifier testing were reused in the WSI test set. To prevent data leakage, these overlapping cases were rigorously excluded from training and validation phases, ensuring independent evaluation at both analytical levels. Furthermore, all samples were sourced from a single institution using standardized protocols; thus, this represents an initial model development phase with controlled technical conditions. Multicenter validation datasets are essential to confirm the generalizability of the models’ performance, particularly across different staining protocols, scanner platforms, and HER2 prevalence distributions. Future work will focus on multi-institutional studies to assess the models’ robustness across varied technical and demographic conditions for clinical deployment. Second, model training and evaluation relied on pathologist annotations as the reference standard. While intra- and inter-observer consistency were assessed, subjective variability in HER2 scoring remains an inherent challenge, especially in borderline categories, such as 1+ and 2+. The accuracy of tumor masking algorithms may also influence models’ output, particularly in WSI-level scoring. Third, although the proposed c-score offers a potential measurement for the granular interpretation of HER2, its clinical utility is yet to be validated in prospective or outcome-linked studies. Whether this c-score can meaningfully inform treatment stratification, especially for HER2-low populations, will require further investigations in larger, annotated cohorts.

5. Conclusions

We developed two transformer-based models for automated HER2 scoring in breast cancer WSIs and demonstrated high concordance with pathologist annotations across multiple tasks. While the foundation model achieved slightly higher accuracy, the lightweight model offered equivalent performance with faster inference. In addition, the proposed continuous c-score captured HER2 expression as a spectrum, offering a promising framework for interpretable, quantitative assessment. These findings support the clinical applicability of AI-assisted HER2 scoring and highlight opportunities for future precision pathology.

Supplementary Materials

The following supporting information can be downloaded at: https://www.mdpi.com/article/10.3390/diagnostics15172131/s1, Figure S1: Training curves of two models; Figure S2: Patches misclassified by models; Table S1: Computational complexity, computational parameters and training hyperparameters of two models.

Author Contributions

Conceptualization, M.-H.C. and H.-H.T.; methodology, M.-H.C.; software, H.-H.T. and C.-J.C.; formal analysis, M.-H.C., H.-H.T. and C.-J.C.; investigation, Y.-H.W.; data curation, H.-H.T. and C.-J.C.; writing—original draft preparation, Y.-H.W. and M.-H.C.; writing—review and editing, Y.-H.W., H.-H.T., C.-J.C. and J.-C.W.; visualization, M.-H.C.; project administration, J.-C.W. All authors have read and agreed to the published version of the manuscript.

Funding

This research was funded by Quanta Medical Technology Foundation (QMTF).

Institutional Review Board Statement

This study received approval from the Research Ethics Review Committee of the Far Eastern Memorial Hospital (No. 113141-E, date of approval: 5 June 2024).

Informed Consent Statement

All blocks used in the clinical diagnosis setting were retrieved from the repository. These slides and images were anonymized and did not contain any personal information. Therefore, the requirement for informed consent was waived.

Data Availability Statement

The original contributions presented in the study are included in the article; further inquiries can be directed to the corresponding author.

Acknowledgments

The authors thank Chu Hsiu-Yi for her support and assistance in writing this article. During the preparation of this manuscript, the authors used GPT-4o for the purposes of making confusion matrices in Figure 3 and Figure 4. The authors have reviewed and edited the output and take full responsibility for the content of this publication.

Conflicts of Interest

Authors Hsin-Hsiu Tsai, Chun-Jui Chien and Jian-Chiao Wang were employed by the company Quanta Computer Inc. The all authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or in the decision to publish the results.

Abbreviations

The following abbreviations are used in this manuscript:
AIArtificial intelligence
H&EHematoxylin and eosin
HER2Human epidermal growth factor 2
IHCImmunohistochemical staining
WSIWhole-slide images

References

  1. Modi, S.; Jacot, W.; Yamashita, T.; Sohn, J.; Vidal, M.; Tokunaga, E.; Tsurutani, J.; Ueno, N.T.; Prat, A.; Chae, Y.S.; et al. Trastuzumab deruxtecan in previously treated HER2-low advanced breast cancer. N. Engl. J. Med. 2022, 387, 9–20. [Google Scholar] [CrossRef]
  2. Hempenius, M.A.; Eenkhoorn, M.A.; Høeg, H.; Dabbs, D.J.; van der Vegt, B.; Sompuram, S.R.; ’t Hart, N.A. Quantitative comparison of immunohistochemical HER2-low detection in an interlaboratory study. Histopathology 2024, 85, 920–928. [Google Scholar] [CrossRef] [PubMed]
  3. Bornstein-Quevedo, L.; de Anda-González, J.; Lara-Torres, C.O.; Flores-Gutiérrez, J.P.; Dorantes-Heredia, R.; Bautista-Piña, V.; Zaragoza-Vargas, P.; Alcaraz-Wong, A.; Soto-Sañudo, A.K.; Mendoza-Ramírez, S.; et al. Navigating HER2-low testing in invasive breast cancer: Update recommendations for pathologists. J. Pers. Med. 2024, 14, 467. [Google Scholar] [CrossRef] [PubMed]
  4. Baez-Navarro, X.; van Bockstal, M.R.; Nawawi, D.; Broeckx, G.; Colpaert, C.; Doebar, S.C.; Hogenes, M.C.H.; Koop, E.; Lambein, K.; Peeters, D.J.E.; et al. Interobserver variation in the assessment of immunohistochemistry expression levels in HER2-negative breast cancer: Can we improve the identification of low levels of HER2 expression by adjusting the criteria? An international interobserver study. Mod. Pathol. 2023, 36, 100009. [Google Scholar] [CrossRef] [PubMed]
  5. Franchina, M.; Pizzimenti, C.; Fiorentino, V.; Martini, M.; Ricciardi, G.R.R.; Silvestris, N.; Ieni, A.; Tuccari, G. Low and ultra-low HER2 in human breast cancer: An effort to define new neoplastic subtypes. Int. J. Mol. Sci. 2023, 24, 12795. [Google Scholar] [CrossRef]
  6. Pham, M.D.; Balezo, G.; Tilmant, C.; Petit, S.; Salmon, I.; Hadj, S.B.; Fick, R.H.J. Interpretable HER2 scoring by evaluating clinical guidelines through a weakly supervised, constrained deep learning approach. Comput. Med. Imaging Graph. 2023, 108, 102261. [Google Scholar] [CrossRef]
  7. Jiang, C.; Li, M.; Zheng, C.; Yan, S.; Kong, L.; Wu, Y.; Zhang, J.; Chao, X.; Cai, X.; Feng, W.; et al. Systematic analysis of factors affecting human epidermal growth factor receptor 2 interpretation consistency: Staining protocols, artificial intelligence-based image standardization, and classification criteria. Lab. Invest. 2025, 105, 104134. [Google Scholar] [CrossRef]
  8. Wróbel, A.; Vandenberghe, M.; Scott, M.; Jones, F.; Matsuo, T.; Boothman, A.M.; Whiteley, J.; Barker, C. Accuracy of human epidermal growth factor receptor 2 (HER2) immunohistochemistry scoring by pathologists in breast cancer, including the HER2-low cutoff: HER2 IHC scoring concordance in breast cancer. Diagn. Pathol. 2025, 20, 35. [Google Scholar] [CrossRef]
  9. Farshid, G.; Armes, J.; Dessauvagie, B.; Gilhotra, A.; Kumar, B.; Mahajan, H.; Millar, E.; Pathmanathan, N.; Snell, C. Independent validation of a HER2-low focused immunohistochemistry scoring system for enhanced pathologist precision and consistency. Mod. Pathol. 2025, 38, 100693. [Google Scholar] [CrossRef]
  10. Shaaban, A.M.; Kaur, T.; Provenzano, E. HER2-low breast cancer-current knowledge and future directions. Medicina 2025, 61, 644. [Google Scholar] [CrossRef]
  11. Rüschoff, J.; Penner, A.; Ellis, I.O.; Hammond, M.E.H.; Lebeau, A.; Osamura, R.Y.; Penault-Llorca, F.; Rojo, F.; Desai, C.; Moh, A.; et al. Global study on the accuracy of human epidermal growth factor receptor 2-low diagnosis in breast cancer. Arch. Pathol. Lab. Med. 2025, 149, 431–438. [Google Scholar] [CrossRef]
  12. Wu, S.; Shang, J.; Li, Z.; Liu, H.; Xu, X.; Zhang, Z.; Wang, Y.; Zhao, M.; Yue, M.; He, J.; et al. Interobserver consistency and diagnostic challenges in HER2-ultralow breast cancer: A multicenter study. ESMO Open 2025, 10, 104127. [Google Scholar] [CrossRef]
  13. Khameneh, F.D.; Razavi, S.; Kamasak, M. Automated segmentation of cell membranes to evaluate HER2 status in whole slide images using a modified deep learning network. Comput. Biol. Med. 2019, 110, 164–174. [Google Scholar] [CrossRef]
  14. Rasmussen, S.A.; Taylor, V.J.; Surette, A.P.; Barnes, P.J.; Bethune, G.C. Using deep learning to predict final HER2 status in invasive breast cancers that are equivocal (2+) by immunohistochemistry. Appl. Immunohistochem. Mol. Morphol. 2022, 30, 668–673. [Google Scholar] [CrossRef]
  15. Priego-Torres, B.M.; Lobato-Delgado, B.; Atienza-Cuevas, L.; Sanchez-Morillo, D. Deep learning-based instance segmentation for the precise automated quantification of digital breast cancer immunohistochemistry images. Expert Syst. Appl. 2022, 193, 116471. [Google Scholar] [CrossRef]
  16. Jha, S.K.; Mishra, P.; Mathur, S.; Singh, G.; Kumar, R.; Aatre, K.; Rengarajan, S. Development and validation of fully automatic deep learning-based algorithms for immunohistochemistry reporting of invasive breast ductal carcinoma. arXiv 2024, arXiv:2406.10893. [Google Scholar] [CrossRef]
  17. Abdulsadig, R.S.; Williams, B.M.; Burlutskiy, N. Leveraging transfer learning and multiple instance learning for HER2 automatic scoring of H&E whole slide images. arXiv 2024, arXiv:2411.05028. [Google Scholar]
  18. Selcuk, S.Y.; Yang, X.; Bai, B.; Zhang, Y.; Li, Y.; Aydin, M.; Unal, A.F.; Gomatam, A.; Guo, Z.; Angus, D.M.; et al. Automated HER2 scoring in breast cancer images using deep learning and pyramid sampling. BME Front. 2024, 5, 0048. [Google Scholar] [CrossRef] [PubMed]
  19. Shamshoian, J.; Shanis, Z.; Cabeen, R.; Yu, L.; Chakraborty, S.; Thibault, M.; Martin, B.; Pagidela, H.; Juyal, D.; Javed, S.A.; et al. Impact of a machine learning-powered algorithm on pathologist HER2 IHC scoring in breast cancer. medRxiv 2025. [Google Scholar] [CrossRef]
  20. Wu, S.; Li, X.; Miao, J.; Xian, D.; Yue, M.; Liu, H.; Fan, S.; Wei, W.; Liu, Y. Artificial intelligence for assisted HER2 immunohistochemistry evaluation of breast cancer: A systematic review and meta-analysis. Pathol. Res. Pract. 2024, 260, 155472. [Google Scholar] [CrossRef]
  21. Thomas, F.; Cormier, B.; Bossard, C.; Poulet, B.; Guinaudeau, E.; Salhi, Y.; Chokri, I.; Jossic, F.; Lambros, L.; De Pinieux, I.; et al. Evaluation of AI-assisted HER2 scoring in breast carcinoma-stained whole-slide images. J. Clin. Oncol. 2024, 42 (Suppl. 16), e13029. [Google Scholar] [CrossRef]
  22. Krishnamurthy, S.; Schnitt, S.J.; Vincent-Salomon, A.; Canas-Marques, R.; Colon, E.; Kantekure, K.; Maklakovski, M.; Finck, W.; Thomassin, J.; Globerson, Y.; et al. Fully automated artificial intelligence solution for human epidermal growth factor receptor 2 immunohistochemistry scoring in breast cancer: A multireader study. JCO Precis. Oncol. 2024, 8, e2400353. [Google Scholar] [CrossRef]
  23. Wu, K.; Zhang, J.; Peng, H.; Liu, M.; Xiao, B.; Fu, J.; Yuan, L. TinyViT: Fast pretraining distillation for small vision transformers. In Computer Vision—ECCV 2022; Avidan, S., Brostow, G., Cissé, M., Farinella, G.M., Hassner, T., Eds.; Springer Nature: Cham, Germany, 2022; pp. 68–85. [Google Scholar] [CrossRef]
  24. Sun, Y.; Zhu, C.; Zheng, S.; Zhang, K.; Sun, L.; Shui, Z.; Zhang, Y.; Li, H.; Yang, L. PathAsst: A generative foundation AI assistant towards artificial general intelligence of pathologyProc. AAAI Conf. Artif. Intel. 2024, 38, 5034–5042. [Google Scholar] [CrossRef]
  25. Lu, M.Y.; Chen, B.; Williamson, D.F.K.; Chen, R.J.; Liang, I.; Ding, T.; Jaume, G.; Odintsov, I.; Le, L.P.; Gerber, G.; et al. A visual-language foundation model for computational pathology. Nat. Med. 2024, 30, 863–874. [Google Scholar] [CrossRef]
  26. Vorontsov, E.; Bozkurt, A.; Casson, A.; Shaikovski, G.; Zelechowski, M.; Severson, K.; Zimmermann, E.; Hall, J.; Tenenholtz, N.; Fusi, N.; et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nat. Med. 2024, 30, 2924–2935. [Google Scholar] [CrossRef] [PubMed]
  27. Ochi, M.; Komura, D.; Ishikawa, S. Pathology foundation models. JMA J. 2025, 8, 121–130. [Google Scholar] [CrossRef]
  28. Wu, C.H.; Chang, M.H.; Tsai, H.H.; Yang, M.L.; Peng, Y.T. Consecutive cytokeratin immunochemistry-supervised algorithm for predicting tumor areas in Ki67 breast cancer images. Hum. Biol. 2024, 94, 729–737. [Google Scholar]
  29. Che, Y.; Ren, F.; Zhang, X.; Cui, L.; Wu, H.; Zhao, Z. Immunohistochemical HER2 recognition and analysis of breast cancer based on deep learning. Diagnostics 2023, 13, 263. [Google Scholar] [CrossRef]
  30. Tarantino, P.; Curigliano, G.; Tolaney, S.M. Navigating the HER2-low paradigm in breast oncology: New standards, future horizons. Cancer Discov. 2022, 12, 2026–2030. [Google Scholar] [CrossRef]
  31. Öttl, M.; Steenpass, J.; Wilm, F.; Qiu, J.; Rübner, M.; Lang-Schwarz, C.; Taverna, C.; Tava, F.; Hartmann, A.; Huebner, H.; et al. Fully automatic HER2 tissue segmentation for interpretable HER2 scoring. J. Pathol. Inform. 2025, 17, 100435. [Google Scholar] [CrossRef]
  32. Gustavson, M.; Haneder, S.; Spitzmueller, A.; Kapil, A.; Schneider, K.; Cecchi, F.; Sridhar, S.; Schmidt, G.; Lakis, S.; Teichert, R.; et al. Abstract PD6-01: Novel approach to HER2 quantification: Digital pathology coupled with AI-based image and data analysis delivers objective and quantitative HER2 expression analysis for enrichment of responders to trastuzumab deruxtecan (T-DXd; DS-8201), specifically in HER2-low patients. Cancer Res. 2021, 81 (Suppl. 4), PD6-01. [Google Scholar] [CrossRef]
  33. Pedraza, A.; Gonzalez, L.; Deniz, O.; Bueno, G. Deep Neural Networks for HER2 Grading of Whole Slide Images with Subclasses Levels. Algorithms 2024, 17, 97. [Google Scholar] [CrossRef]
  34. Wang, J.; Zhu, X.; Chen, K.; Hao, L.; Liu, Y. HAHNet: A convolutional neural network for HER2 status classification of breast cancer. BMC Bioinform. 2023, 24, 353. [Google Scholar] [CrossRef] [PubMed]
Diagnostics 15 02131 g001
Figure 1. Overview of the HER2 scoring workflow. Two transformer-based models (Virchow and TinyViT) were trained using pathologist-annotated patches to classify HER2 immunoreactivity into five categories (0, 1+, 2+, 3+, and non-tumor. Red means different from previous labeling). After patch-level training, the models were applied to WSIs to generate HER2 predictions and heatmaps of tumor regions. The WSI-level score was estimated based on the percentage of predicted patch classes, following ASCO/CAP guidelines. Abbreviations: HER2 = human epidermal growth factor 2; WSI = whole-slide image.
Figure 1. Overview of the HER2 scoring workflow. Two transformer-based models (Virchow and TinyViT) were trained using pathologist-annotated patches to classify HER2 immunoreactivity into five categories (0, 1+, 2+, 3+, and non-tumor. Red means different from previous labeling). After patch-level training, the models were applied to WSIs to generate HER2 predictions and heatmaps of tumor regions. The WSI-level score was estimated based on the percentage of predicted patch classes, following ASCO/CAP guidelines. Abbreviations: HER2 = human epidermal growth factor 2; WSI = whole-slide image.
Diagnostics 15 02131 g001
Diagnostics 15 02131 g002
Figure 2. Confusion matrices for patch-level HER2 classification. (A) Intra-observer comparison (Dr. A initial labeling vs. Dr. A relabeling). (B) Inter-observer comparison (Dr. A vs. Dr. B). (C) Dr. A vs. Model V. (D) Dr. A vs. Model T. Each matrix shows row-normalized percentages by ground truth category. Abbreviation: HER2 = human epidermal growth factor 2.
Figure 2. Confusion matrices for patch-level HER2 classification. (A) Intra-observer comparison (Dr. A initial labeling vs. Dr. A relabeling). (B) Inter-observer comparison (Dr. A vs. Dr. B). (C) Dr. A vs. Model V. (D) Dr. A vs. Model T. Each matrix shows row-normalized percentages by ground truth category. Abbreviation: HER2 = human epidermal growth factor 2.
Diagnostics 15 02131 g002
Diagnostics 15 02131 g003
Figure 3. Confusion matrices and HER2 heatmaps of the two models for WSI-level HER2 scoring. (A) Model V vs. ground truth. (B) Model T vs. ground truth. Ground truth was according to original reports. (C) Heatmap examples of Model V (left) and Model T (right). Model V misclassified more of the non-tumor patches as 0 or 1+ (blue arrows). Abbreviations: HER2 = human epidermal growth factor 2; WSI = whole-slide image.
Figure 3. Confusion matrices and HER2 heatmaps of the two models for WSI-level HER2 scoring. (A) Model V vs. ground truth. (B) Model T vs. ground truth. Ground truth was according to original reports. (C) Heatmap examples of Model V (left) and Model T (right). Model V misclassified more of the non-tumor patches as 0 or 1+ (blue arrows). Abbreviations: HER2 = human epidermal growth factor 2; WSI = whole-slide image.
Diagnostics 15 02131 g003
Diagnostics 15 02131 g004
Figure 4. Correlation of c-scores between Model V and Model T. (A) Scatter plot showing the correlation between c-scores generated by Model V (x-axis) and Model T (y-axis) across all test cases (n = 66). Each dot represents a single case, with colors indicating HER2 IHC ground truth: red for 3+, blue for 2+, light green for 1+, and gray for 0. (B) Zoomed-in view of the same scatter plot, focusing on the c-score range 0−0.5. Abbreviations: HER2 = human epidermal growth factor 2; IHC = immunohistochemistry.
Figure 4. Correlation of c-scores between Model V and Model T. (A) Scatter plot showing the correlation between c-scores generated by Model V (x-axis) and Model T (y-axis) across all test cases (n = 66). Each dot represents a single case, with colors indicating HER2 IHC ground truth: red for 3+, blue for 2+, light green for 1+, and gray for 0. (B) Zoomed-in view of the same scatter plot, focusing on the c-score range 0−0.5. Abbreviations: HER2 = human epidermal growth factor 2; IHC = immunohistochemistry.
Diagnostics 15 02131 g004
Table 1. Case details.
Table 1. Case details.
HER2: 0HER2: 1+HER2: 2+HER2: 3+Total
Dataset131812649
Training51310331
Validation541111
Test31127
WSI test set1519181466 *
* Includes 7 WSIs in test set. To prevent data leakage, the 7 overlapping slides between patch and WSI test sets were completely excluded from all training and validation phases, serving solely as independent test cases for both patch-level and WSI-level evaluation. Abbreviations: HER2 = human epidermal growth factor 2; WSI = whole-slide image.
Table 2. Number of patches for model training.
Table 2. Number of patches for model training.
Non-Tumor01+2+3+Total
Training986794534282912687
Validation5412451751583231442
Test273181165158203980
Total180012208745986175109
Table 3. Model performance in three clinical binary tasks.
Table 3. Model performance in three clinical binary tasks.
Binary TasksModel V (Virchow)Model T (TinyVit)
AccuracyPrecisionSensitivityF1 ScoreAccuracyPrecisionSensitivityF1 Score
0, 1+, 2+ vs. 3+ *100%100%100%1.0095.5%82.4%100%1.00
0, 1+ vs. 2+ **84.6%100%55.6%0.71480.8%83.3%55.6%0.667
0 vs. 1+, 2+ ***94.2%97.2%94.6%0.95986.5%94.1%86.5%0.901
* Includes all 66 test cases, corresponding to selecting patients eligible for anti-HER2 therapy. ** Excludes 3+ cases, corresponding to select candidates for ISH test. *** Represents HER2-low vs. HER2-negative, based on the ASCO/CAP and emerging ADC treatment criteria. Abbreviations: ADC = antibody–drug conjugate; HER2 = human epidermal growth factor 2; ISH = in situ hybridization.
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

Wang, Y.-H.; Chang, M.-H.; Tsai, H.-H.; Chien, C.-J.; Wang, J.-C. Transformer-Based HER2 Scoring in Breast Cancer: Comparative Performance of a Foundation and a Lightweight Model. Diagnostics 2025, 15, 2131. https://doi.org/10.3390/diagnostics15172131

AMA Style

Wang Y-H, Chang M-H, Tsai H-H, Chien C-J, Wang J-C. Transformer-Based HER2 Scoring in Breast Cancer: Comparative Performance of a Foundation and a Lightweight Model. Diagnostics. 2025; 15(17):2131. https://doi.org/10.3390/diagnostics15172131

Chicago/Turabian Style

Wang, Yeh-Han, Min-Hsiang Chang, Hsin-Hsiu Tsai, Chun-Jui Chien, and Jian-Chiao Wang. 2025. "Transformer-Based HER2 Scoring in Breast Cancer: Comparative Performance of a Foundation and a Lightweight Model" Diagnostics 15, no. 17: 2131. https://doi.org/10.3390/diagnostics15172131

APA Style

Wang, Y.-H., Chang, M.-H., Tsai, H.-H., Chien, C.-J., & Wang, J.-C. (2025). Transformer-Based HER2 Scoring in Breast Cancer: Comparative Performance of a Foundation and a Lightweight Model. Diagnostics, 15(17), 2131. https://doi.org/10.3390/diagnostics15172131

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