DTCA ‐Net: A Dilated Transformer Channel Attention With External Attention‐Based Network for Oral Cancer Segmentation

ABSTRACT

Accurate segmentation of oral cancer in histopathological images is essential for early diagnosis and treatment. While U‐Net‐based architectures have shown promise, balancing high accuracy with low computational complexity remains a key challenge. This work aims to design an efficient model that enhances feature extraction without significantly increasing model size. We propose DTCA‐Net, an enhanced U‐Net variant that integrates three key modules: (1) a Dilated Convolution and Transformer Channel Attention (DTCA) block to capture global and local features, (2) an Inverted External Attention (IEA) block to model intersample relationships and dataset‐level context, and (3) an Attention Bridge (AB) module connecting encoder and decoder to improve multiscale feature fusion. The model was evaluated on two public datasets: Oral Cavity‐Derived Cancer (OCDC) and ORal Cancer Annotated (ORCA). DTCA‐Net achieved strong performance, with a Dice score of 91.45% and IoU of 84.44% on the OCDC dataset. On the ORCA dataset, it obtained a Dice score of 83.66% and IoU of 71.96%, surpassing existing baselines in both accuracy and efficiency. We effectively combine lightweight architecture with rich feature extraction capabilities, making DTCA‐Net a promising tool for automatic oral cancer segmentation in clinical workflows. The entire codes of our model are publicly available through this Github Link: https://github.com/shivamsingh‐gpu/dtca‐net .