Multi-class semantic segmentation of oral cancer tissues on whole-slide images: A SuperPixel enhanced deep learning approach

Oral cancer remains a significant global health challenge, where early and accurate histopathological diagnosis is essential for improving patient outcomes. Whereas deep learning has shown promise in assisting pathologists, current methods face significant challenges by the need for exhaustive manual annotations, class imbalance, and limited contextual insight from patch-based analyses. This work introduces a deep learning method for the automated semantic segmentation of 10 different tissue types within oral cancer tumor regions of interest (ROIs) on whole-slide images (WSIs). To reduce annotation burden and enhance precision, a superpixel-based approach is proposed for data annotation and incorporated a superpixel refinement step during inference to improve boundary delineation while reducing noisy pixels. The model was rigorously evaluated across nine combinations of architectures and backbones, using 444.038 patches from 10 different patients for training and validation (via a 5-fold Leave-One-Fold-Out Group K-Fold Cross-Validation) and 78.103 patches from two independent patients for testing. The present approach achieved a robust average accuracy of 0.921 and micro Intersection over Union (IoU) of 0.859, demonstrating robust results even in under-represented and morphologically diverse tissue classes. Despite challenges related to class imbalance and local context limitations, the proposed methodology consistently provided high-quality predictions. Furthermore, seamless integration into the open-source QuPath platform enables WSI ROI reconstruction, highlighting its potential for real-world clinical deployment in oral cancer diagnostics.