A comparative study of vision transformer architectures evaluating MYC status prediction in diffuse large B cell lymphoma

Predicting MYC gene rearrangements in diffuse large B-cell lymphoma (DLBCL) from Hematoxylin and Eosin (H&E) images can optimize diagnostic triaging. This is crucial because Fluorescence in Situ Hybridization (FISH) testing remains costly and time-consuming. We evaluated five optimized Vision Transformer (ViT) paradigms to benchmark digital prescreening: Basic ViT, T2T-ViT, CLIP-ViT, Swin Transformer, and Hierarchical ViT. Models were tested on 463,200 normalized tissue patches from 120 whole-slide images (WSIs). Architectures were assessed on predictive accuracy, Area Under the Receiver Operating Characteristic (AUROC), and inference throughput. Hierarchical ViT achieved the highest absolute classification accuracy through multi-scale feature modeling. Among end-to-end models, Swin Transformer offered a favorable AUROC-throughput trade-off. It operated nearly three times faster than Hierarchical ViT at 5.16 vs. 1.78 batches/s. Meanwhile, T2T-ViT achieved the fastest overall inference at 11.17 batches/s. Grad-CAM visualizations showed top models focused on morphologically relevant tumor regions. However, visual heatmaps do not confirm biological correctness. Overall, multi-scale ViT architectures effectively predict patch-level MYC status in DLBCL, with Swin Transformer balancing speed and performance. Because this study uses a single-center, patch-level dataset, multicenter and patient-level WSI validations remain necessary before establishing clinical utility.