Keratin-net: a lightweight self supervised fusion framework for simultaneous classification and localization of keratinization in oral cancer histopathology

Keratinization is a critical histopathological feature for grading oral squamous cell carcinoma. Therefore, to automated its quantification, we propose Keratin-Net, a lightweight self-supervised stain-fusion framework that performs simultaneous classification and fine-grained localization of keratinized regions, eliminating the need for explicit segmentation or pixel-level annotations. Keratin-Net integrates dual-feature backbones, a trainable H&E fusion module, and stain-specific generative SSL pretraining tasks ( HE \(_{rgb}\) 2Eos and HE \(_{rgb}\) 2Hem ) to learn stain-aware embeddings, producing class-specific gradients for Grad-CAM–based fine-grained localization. Benchmarking against stain-specific SSL baselines (RGB2H, RGB2RGB, P \(_{stain}\) ), supervised ResNet-50, and Pathology Foundation models (UNI, Virchow2, H-Optimus-1, CONCH), Keratin-Net achieved in-distribution AUC of 0.99 and F1 of 94.48 ± 0.01%, and maintained strong out-of-distribution performance (F1 = 92.48 ± 1.34%). Patch-level localization reached IoU 0.597 and Dice 0.732, surpassing all baselines. Despite its compact design (1.18M parameters, 7 MB memory), Keratin-Net ranked first in MCDA for both classification and localization, balancing accuracy, generalization, and efficiency. The results highlight Keratin-Net’s ability to perform fine-grained localization without requiring pixel-level manual annotations, underscoring the importance of the proposed domain-specific Self supervised pretraining tasks and architecture design. This approach can be applied to other cancer types, such as skin and esophageal cancers, where keratin identification also serves as a key pathological marker, showcasing its potential for broader clinical applications.