DeepPatchNet: A deep learning model for enhanced screening and diagnosis of oral cancer

Oral cancer remains a serious global health challenge, significantly affecting patient survival and quality of life. While convolutional neural networks (CNNs) have historically dominated image classification tasks, recent advances suggest that transformer-based models may offer superior performance—albeit with high data and computational demands. In this study, we present DeepPatchNet, a novel deep learning architecture that integrates DeepLabV3+ and ConvMixer to address these limitations. Designed for histopathological image classification, DeepPatchNet provides a lightweight, interpretable, and high-performing solution. We evaluated the model on the NDB-UFES dataset (3763 images) and an independent H&E-stained OSCC dataset (1020 images), benchmarking its performance against state-of-the-art models including Vision Transformers (ViTs)[1], [2], InceptionResNetV2, VGG19, and ConvNeXt. DeepPatchNet achieved superior performance with 86.71% accuracy, 86.80% precision, 86.71% recall, and an F1 score of 86.75%, outperforming all comparison models. Furthermore, the integration of Gradient-weighted Class Activation Mapping (Grad-CAM) enhances interpretability by visually highlighting diagnostically relevant features, addressing a key barrier to clinical adoption. While our results are promising, further validation in real-world clinical settings is needed. DeepPatchNet shows strong potential as a reliable decision support tool for early oral cancer detection and diagnosis.