Application Research Of Swin Transformer And Convolutional Neural Network In Dermoscopic Image Segmentation

Skin cancer is one of the three major cancers,and malignant melanoma is the most serious type of skin cancer.Early detection and clinical intervention can increase survival rates to 95%,compared to 15% for patients with advanced disease,a paradox that highlights the importance of early diagnosis of skin cancer.The early diagnosis of skin cancer mainly relies on dermatologists to mark the lesion area in dermoscopic images before diagnosis,which is not only time-consuming and laborintensive,but also the diagnosis result is subjective and irreproducible.Hence,there is an urgent need to propose a reliable computer-aided diagnostic method to alleviate the limitations of manual analysis of dermoscopic images.Deep learning has made breakthroughs with the upgrading of hardware,and more deep learning methods are focusing on medical image segmentation.Convolutional Neural Networks and Transformer become mainstream in the medical image segmentation field.Based on Convolutional Neural Networks and Transformer,the main research of the paper includes:(1)A novel U-shaped network named PCF-Net is proposed for the accurate segmentation of lesion areas in dermoscopic images.The baseline network of PCF-Net is UNet,which consists of encoder,decoder,bottleneck and skip connections.Three schemes are designed to optimize the UNet baseline network.1)Explore the introduction of a multi-scale grouped dilated convolutional feature extraction module in the bottleneck to extract features with different receptive fields to ensure that the network has the input of multiscale features.2)Explore the problem of embedding the global context information complementary module in the skip connection to compensate for the loss of feature information due to successive downsampling,and extract the global information of a large receptive field by Involution.3)Explore the effect of different attention mechanisms in the decoder on the network segmentation performance,and design a two-branch fusion attention module to enhance the ability of the network to highlight the lesion region and effectively retain the boundary information of the lesion region.(2)To alleviate the problem of the poor ability of convolutional neural networks to model long-range dependencies caused by convolutional windowing,a multi-encoder network called iU-Net hybrid of Swin Transformer and convolutional neural network is designed.The encoder part of iU-Net consists of a primary encoder and a secondary encoder.The local feature information(e.g.,the boundary of the lesion region)captured by the sub-encoder is used as information complementary to the global context information(e.g.,the contour information of the lesion region)extracted by the primary encoder.To effectively fuse the feature information from different encoders,a wave function-based feature fusion module is proposed.This module represents the image blocks as discrete waves and operates feature fusion in the complex domain in a vector manner.In Addition,this paper studies the effect of three different encoder structures on the feature maps of the decoding stage.(3)The segmentation performance of the proposed method is verified by extensive experiments based on two dermoscopic image datasets,ISIC2017 and ISIC2018.On the ISIC2018 and PH2 datasets,the paper adopts a "cross-validation" approach,using the model trained on the ISIC2018 dataset to test on the full PH2 dataset to verify the cross-dataset capability of the model.Comparative experiments on the segmentation performance of PCF-Net and iU-Net networks are done on the ISIC2017 dataset and evaluated by Dice coefficient,IoU and other metrics.The generalization ability of the proposed network was verified using the joint Montgomery,JRST,and NIH lung region datasets,and the experimental results showed that the proposed method has excellent generalization ability.