Tiny-Lesion Segmentation in OCT via Multi-scale Wavelet Enhanced Transformer

The accurate segmentation of retinal lesions from OCT images can greatly aid ophthalmologists in evaluating retinal diseases. However, it remains a challenge to accurately segment retinal lesions in OCT images. This is due to the complicated pathological features of retinal diseases, resulting in severe regional scale imbalance between different lesions, and leading to the problem of target tendency of the network during training, subsequently resulting in the segmentation performance reduction for tiny-lesion. Aiming to solve these challenges, we propose a novel multi-scale wavelet enhanced transformer network for tiny-lesion segmentation in retinal OCT images. In the proposed model, we first design a novel adaptive wavelet down-sampling module combined with the pre-trained ResNet blocks as the feature extractor network, which can generate a wavelet representation to improve the model's interpretability while avoiding feature loss, and further enhancing the ability of the network to represent local detailed features. Meanwhile, we also develop a novel multi-scale transformer module to further improve the model's capacity of extracting the multi-scale long-dependent global features of the retinal lesions in OCT images. Finally, the proposed method is evaluated on the public database of AI-Challenge 2018 for retinal edema lesions joint segmentation, and the results indicate that the proposed method achieves better segmentation performance than other state-of-the-art networks, especially for tiny PED lesions with very small regional proportions.