Abstract 2431: Enhanced cancer detection using TransUnet for low-resolution histopathology images across multiple cancer types
Muhammad Shaiq Paracha, Faisal F. Khan, Arsalan Riaz, Madina Shirdel · Cancer Research · 2025
Abstract Digital histopathology has become an indispensable tool for cancer diagnosis and prognosis, enabling pathologists to analyze tumor morphology using Whole Slide Images (WSI). However, the high cost and limited availability of high-resolution imaging equipment in underdeveloped regions pose significant challenges. This study investigates the use of Low-Cost Low-Resolution (LCLR) histopathology images for detecting four cancer sites: Oral, Gastrointestinal, Colorectal and Breast cancers. We compared the performance of four models: AlexNet, EfficientNet, Vision Transformer and TransUnet. These models were trained on two datasets: the Tabassum et al. (2020) OSCC dataset (1, 223 images, 10X and 40X magnification) used for selecting the right model and HistoVault v1 (17, 500 images, 10X and 40X magnification) for training the selected model for all cancer sites. TransUnet, which integrates U-Net's encoder-decoder architecture with Transformers, was fine-tuned with hyperparameters including a learning rate of 1e-5, batch size of 32 and a dropout rate of 20%. The evaluation metrics included accuracy, sparse categorical cross-entropy loss and confusion matrices to quantify the models' robustness. TransUnet achieved the highest accuracy of 96.8% with a loss of 3% over other models on Tabassum et al. (2020) OSCC dataset. Individual accuracies for Oral, Gastrointestinal, Colorectal and Breast cancers were 98.7%, 95.9%, 96.1% and 97.2% on Histovault v1 dataset. Its coarse-to-fine attention refinement enabled precise segmentation in noisy, low-resolution images, outperforming other models. Notably, the model exhibited robust generalization for multi-class classification, achieving a test accuracy of 90.5% in differentiating eight classes (Non Cancerous and Cancerous with each cancer site). TransUnet possesses significant promises for LCLR image-based cancer detection, especially in environments with limited resources. Its hybrid design improves accuracy and segmentation precision in challenging datasets by combining the advantages of transformers and convolutional neural networks. With the objective to improve precision diagnosis, future research will investigate cancer grading and use explainable AI algorithms. Citation Format: Muhammad Shaiq Paracha, Faisal F. Khan, Arsalan Riaz, Madina Shirdel. Enhanced cancer detection using TransUnet for low-resolution histopathology images across multiple cancer types [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 2431.