Multi-scale swin transformer enabled automatic detection and segmentation of lung metastases using CT images

Detection of lung nodules has always been a challenging topic in medical diagnosis and requires a lot of manual work. Recently, the outgrowth of the convolutional neural network (CNN) draws a lot of attention for its strong and robust ability to address detection and segmentation problems in medical images. In this paper, we present a modified Swin Transformer model (ST) integrated with a novel Multi-Scale Multi-Level Feature Fusion Reorganization (MSMLFFR) Module. We constructed a modified Swin Transformer network having a Local Branch and a Global Branch to consider both global and local features from the CT images and manual annotations. Our proposed model captures global and local dependencies in image feature learning and provides wide-ranging contextual information. Novel MSMLFFR module designed for skip connection optimization fuses the global and local features iteratively throughout multiple transformers. Our model results have outperformed the state-of-the-art method Unet and Swin Transformer in lung lesion segmentation on the LIDC-IDRI. Our model achieved 96% DSC, whereas Unet achieved 82% DSC and Swin Transformer achieved 88% DSC on LIDC-IDRI dataset. Proposed model achieved 89.7% MCC score indicating high correlation between the ground truth and predicted segmentation of lung metastases. Extensive performance evaluation of our model ST-MSMLFFR on five benchmark datasets namely ILD, ANODE09, LIDC-IDRI, ELCAP, and LUNA16, shows the potential of our model for lesion detection and segmentation.