DETECTING ABNORMAL AREAS ON BRAIN MRI IMAGES WITH SWIN-UNET

Abstract

To identify abnormal areas on brain MRI images, radiologists need to examine many slices from the image set. This research helps automatically suggest abnormal areas of the brain on MRI images. The Unet, ResNet, Swin-Unet models are trained on the Can Tho University of Medicine and Pharmacy Hospital data set combined with the LGG data set to segment images with or without abnormal regions. The model will then suggest the abnormal region through the boundary drawn around it. Experimental results show that, when dividing random data by image, the Swin-Unet model achieves the highest accuracy with 0.88, along with Recall, Precision and F1 Score of 0.96, 0.71, and 0.82 respectively. For determining the location and shape of the abnormal region, Swin-Unet also demonstrated high performance with mIoU reaching 0.89 and mDSC reaching 0.91. When dividing the data by patient, the Swin-Unet model once again showed good performance with Accuracy reaching 0.86, along with Recall of 0.88, Precision of 0.79, F1 Score of 0.83, and for mIoU it achieved 0.84 and mDSC reached 0.89. Research results show that the Swin-Unet model has good results in the problem of detecting abnormal areas on brain MRI images.