Technical Summary: Efficiency and Interpretability in MYC Status Prediction Using Vision Transformers

Problem Statement

Predicting the status of the MYC proto-oncogene in Diffuse Large B-Cell Lymphoma (DLBCL) is critical for risk stratification and treatment planning, as MYC-positive cases are associated with aggressive disease progression. Currently, the gold standard for detection is Fluorescence In Situ Hybridization (FISH), a resource-intensive, expensive, and time-consuming laboratory test. While deep learning offers the potential to predict biomarker status from standard Hematoxylin and Eosin (H&E) whole-slide images (WSIs), existing Convolutional Neural Networks (CNNs) often struggle with the multi-scale structural complexity of lymphoma tissue, specifically their inability to capture long-range spatial dependencies. Furthermore, there is a lack of systematic, head-to-head benchmarks comparing distinct Vision Transformer (ViT) architectures regarding their trade-offs between predictive accuracy, computational efficiency, and biological interpretability in the context of DLBCL.

Methodology

The study established a comprehensive benchmark evaluating five structurally diverse Vision Transformer architectures on a large-scale dataset of digital pathology images.

• Dataset: The research utilized 120 high-resolution H&E-stained WSIs from 120 distinct patients (60 MYC-positive and 60 MYC-negative) obtained from Hospital Universiti Sains Malaysia. The dataset was strictly partitioned at the patient level to prevent data leakage, resulting in a training set of 84 WSIs, a validation set of 12 WSIs, and a testing set of 12 WSIs.
• Data Preprocessing:
  • Compression: Lossless compression was applied to preserve high-frequency morphological details (e.g., nuclear chromatin texture).
  • Normalization: Macenko normalization was used to correct color variations across slides.
  • Patch Extraction: WSIs were tiled into non-overlapping 256×256 pixel patches. An automated low-entropy filtration process removed background and low-cellularity regions, yielding a final dataset of 463,200 high-quality patches (231,600 per class).
• Architectures Evaluated:
  1. Basic ViT: A standard global self-attention model to evaluate pure global context.
  2. Tokens-to-Token (T2T) ViT: Incorporates recursive token aggregation to capture local structural information.
  3. CLIP-ViT B/16: A pre-trained foundation model (frozen backbone, fine-tuned head) to test feature transferability from natural images.
  4. Swin Transformer: Utilizes a shifted-window mechanism for hierarchical, local attention with linear computational complexity.
  5. Hierarchical ViT (PVT-v2): A pyramid-based architecture mimicking multi-resolution feature extraction to handle both subcellular and global tissue patterns.
• Training Protocol: All models were trained for 100 epochs on an Nvidia Tesla T4 GPU using PyTorch. A uniform hyperparameter configuration was applied (AdamW optimizer, learning rate $1\times10^{-4}$, weight decay 0.05, batch size 32, Cosine Annealing scheduler) to ensure a fair comparison.
• Evaluation Metrics: Performance was assessed using Accuracy, Precision, Recall, F1-score, Area Under the Receiver Operating Characteristic (AUROC), and Average Precision (AP). Computational efficiency was measured as inference throughput (iterations per second, it/s).
• Statistical Validation: 95% confidence intervals were calculated using Wald's asymptotic formula. Pairwise comparisons of AUROC curves were conducted using DeLong's test with Holm-Bonferroni correction.
• Explainability: Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to visualize the specific tissue regions influencing model predictions.

Key Results

The study revealed distinct trade-offs between diagnostic accuracy and computational efficiency among the architectures:

• Diagnostic Accuracy: The Hierarchical ViT achieved the highest overall classification accuracy (0.82) and the most balanced F1-scores across both classes. It demonstrated superior ability to avoid false negatives and false positives compared to other models.
• Risk Stratification (AUROC): The Swin Transformer achieved the highest AUROC (0.90), statistically significantly outperforming the Hierarchical ViT (0.89, $p=0.040$) and all other architectures. This indicates the Swin Transformer is highly effective at ranking risk probabilities, making it suitable for continuous screening.
• Computational Efficiency: A clear inverse relationship was observed between accuracy and speed. The CLIP-ViT was the fastest (14.53 it/s) but had the lowest accuracy (0.69). The Hierarchical ViT was the slowest (1.78 it/s), requiring approximately 36 minutes per slide. The Swin Transformer offered a middle ground (5.16 it/s, ~12.5 minutes per slide), delivering nearly three times the throughput of the Hierarchical ViT.
• Interpretability: Grad-CAM visualizations confirmed that high-performing models (Hierarchical ViT and Swin Transformer) focused on morphologically relevant tumor regions, such as pleomorphic nuclei and hyperchromatic chromatin. In contrast, lower-performing models (Basic ViT) often attended to background stroma or artifacts.

Significance and Claims

The paper claims to provide the first comprehensive, head-to-head benchmark of five distinct transformer paradigms for MYC status prediction in DLBCL. Its primary contributions are:

1. Structural Blueprint: It establishes a structural blueprint for digital pathology workflows under single-center and patch-level constraints, proving that no single architecture is universally superior.
2. Trade-off Identification: The study explicitly identifies the "Pareto-optimal" solutions for different clinical needs:
   • The Hierarchical ViT is identified as the optimal choice for definitive diagnostic classification where maximum accuracy is required, despite higher computational costs.
   • The Swin Transformer is identified as the optimal choice for risk stratification and high-throughput screening, offering a statistically significant edge in AUROC and nearly triple the inference speed of the most accurate model.
3. Clinical Trust: By integrating Grad-CAM, the study demonstrates that high-performing transformers ground their predictions in biologically relevant features (nuclear morphology), bridging the gap between "black-box" metrics and clinical utility.

The authors maintain a modest stance, noting that these findings are based on a single-center cohort and patch-level analysis. They assert that while the results provide a robust framework for automated screening, real-world clinical deployment requires multi-center validation and the integration of slide-level aggregation methods (e.g., Multiple Instance Learning) to translate patch-level predictions into final patient diagnoses.