[Topology-aware self-attention network for automatic segmentation of oral potentially malignant disorder images].

Objective

To develop an automatic oral lesion segmentation model (TopoFormer) that integrates topological data analysis (TDA) with a Transformer architecture, thereby improving the accuracy and robustness of lesion segmentation for oral potentially malignant disorders (OPMD) and oral squamous cell carcinoma (OSCC).

Methods

Experiments were conducted on a public comprehensive oral cavity image dataset released by Piyarathne et al., containing 2 271 oral white-light photographs from 623 patients and covering three categories: At the patient level, the data were stratified and randomly split into training, validation, and test sets with a ratio of 0.8/0.1/0.1. SegFormer was used as the backbone network, and a self-designed topological attention module (TAM) was embedded in the deep feature space. TAM leveraged persistent homology to extract robust topological structures from feature maps and dynamically suppressed low-persistence noise. Performance was evaluated using the Dice similarity coefficient (DSC), mean intersection over union (mIoU), mean pixel accuracy (mPA), 95% Hausdorff distance (HD95), and Betti number error (BNE), and was compared with mainstream segmentation networks (U-Net, DeepLabV3+, and SegFormer).

Results

On the test set, TopoFormer achieved a DSC of 0.801, an mIoU of 71.1%, an mPA of 82.31%, reduced HD95 to 9.62 pixels, and obtained a BNE of 1.28. Compared with the baseline model, the proposed model effectively reduced false positives caused by specular reflections and produced smoother and more continuous lesion boundaries.

Conclusions

TopoFormer demonstrates higher accuracy and robustness in segmenting OPMD and OSCC lesions. Topological priors can enhance segmentation performance in complex oral environments, indicating use potential in computer-aided diagnosis.