PDTANet: a context-guided and attention-aware deep learning method for tumor segmentation of guinea pig colorectal OCT images

Abstract

Optical coherence tomography (OCT) technology has significant potential value in the application of early gastrointestinal tumor screening and intraoperative guidance. In the application of diagnosing gastrointestinal diseases, a key step of OCT image intelligent analysis system is to segment the tissues and layers accurately. In this paper, we propose a new encoder-decoder network named PDTANet, which contains a global context-guided PDFF module and a lightweight attention-aware triplet attention (TA) mechanism. Moreover, during the model training stage, we adopt a region-aware and boundary-aware hybrid loss function to learn and update model parameters. The proposed PDTANet model has been applied for automatic tumor segmentation of guinea pig colorectal OCT images. The experimental results show that our proposed PDTANet model has the ability to focus on and connect global context and important feature information for OCT images. Compared with the prediction results of the model trained by the traditional Unet model and Dice loss function, the PDTANet model and a combination of dice and boundary related loss function proposed as the hybrid loss function proposed in this paper have significantly improved the accuracy of the segmentation of tissue boundaries, especially the surface Dice metric, which is improved by about 3%.