A Robust Deep Learning Framework For Multi-Resolution Edge Detection in MR Images

Abstract

The analysis of anatomical structure and clinical decision-making depends on the accurate detection of edges of anatomical boundaries by using magnetic resonance (MR) imaging. Luckily, because of noise, low contrast, and inadequate multi-resolution representation in the image, traditional methods of edge detection and even deep learning-based traditional approaches have been unable to give satisfactory results. A multi-resolution deep learning approach is proposed for edge detection of anatomical structures using MR imaging. It will overcome the challenges posed by traditional edge detection methods and will focus on improving the edge-aware feature representation using transformers, edge attention, and multi-scale feature extraction. For efficient fusion of the spatial and semantic features and stabilization of the multi-level learning process, deep supervision and strengthened skip connections are introduced. In addition, in order to optimize the performance of pixels, boundaries, and structures simultaneously, a hybrid loss function is proposed. The proposed approach surpasses both the traditional approaches for edge detection and modern deep learning frameworks in respect to entropy, MSE, PSNR, SSIM, EKI, and FSIM metrics through extensive testing on MR image database. Therefore, the experimental results provide the accuracy of 96.7%, specificity of 95.8%, and F1 score of 95.3%, which is higher than the state-of-the-art methods that increase the effectiveness of the proposed deep learning technique.