Edge of discovery: Enhancing breast tumor MRI analysis with boundary-driven deep learning

Manually segmenting breast lesion images poses a labor-intensive and expensive undertaking for radiologists. Therefore, the adoption of an automated diagnostic approach becomes imperative, aiming to precisely segment breast lesions and mitigate the associated challenges. The segmentation of malignant regions is a crucial process in the comprehensive examination of breast Magnetic Resonance Imaging (MRI). However, achieving proficient segmentation through the effective utilization of deep learning methodologies continues to pose a formidable challenge. The complexity of the task is underscored by the substantial variations evident in the dimensions, structural attributes, and visual characteristics inherent to diverse malignancy types. This research proposes an Enhance Edge U-Net (EEU-Net) model to address these concerns. In our proposed model, a Deep Convolutional Neural Network (DCNN) is employed that follows an architecture comprising both an encoder and a decoder, drawing inspiration from the U-Net framework. The utilization of the EEU-Net model facilitates enhanced tumor localization through the integration of MRI data associated with boundaries with the primary data obtained from breast MRI. In the decoding stage, the information pertaining to boundaries is derived from source MRI scans with diverse scales combined with the relevant nearby surrounding information. To enhance the performance of this segmentation model, a loss function is incorporated. The incorporation of boundary information in this loss function enhances the learning process, resulting in more precise outcomes. The results of our experiments indicate that EEU-Net is superior in efficiency and reliability pertaining to the segmentation of breast tumors as compared to other techniques. It outperforms alternative segmentation methods in terms of the Jaccard and Dice coefficients when evaluated on the publicly accessible RIDER breast cancer MRI dataset. Specifically, EEU-Net achieved Jaccard and Dice coefficients of 0.804(±0.024) and 0.884(±0.017), respectively, on the RIDER breast MRI datasets. Additionally, the statistical analysis for the proposed model, utilizing a paired t-test, uncovered a noteworthy difference with a p-value less than 0.05.