FTSEGNET: A Novel Transformer-Based Fundus Tumor Segmentation Model Guided by Pre-Trained Classification Results

Fundus tumors are the most severe retinopathies, and the deep learning segmentation model can locate the lesions and segment the lesion contour as a computer-aided diagnosis method. However, the existing algorithms for segmenting fundus tumor images have poor performance, making them unsuitable for practical clinical use. This project addresses the major issues of data and performance limitations in fundus tumor image segmentation tasks. We collect a new dataset for fundus tumor image segmentation, named FTS, which contains 254 pairs of fundus images with fundus tumor lesions and their segmentation reference images. Furthermore, a new fundus tumor segmentation network called FTSegNet is proposed in this paper. The key component in FTSegNet is the Classification Prior Block(CPB), which can provide the prior feature from classification pre-trained and guide segmentation. To better extract feature information, the Transformer and convolutional layers have been effectively combined. Qualitative and quantitative experiments are conducted on the FTS dataset to verify the effectiveness of the proposed model. We also explore the effectiveness of the CPB and different loss functions in FTSegNet. This method can provide methodological concepts for future fundus tumor segmentation tasks.