MSR-Net: Multi-Scale Residual Network Based on Attention Mechanism for Pituitary Adenoma MRI Image Segmentation

Accurate segmentation of pituitary adenoma lesions is essential for effective diagnosis and treatment planning. However, traditional algorithms struggle with this task due to the variability in lesion shape, position, and the presence of extensive background areas. In response to the above challenges, we develop a segmentation framework for pituitary adenoma that integrates multi-scale residual, channel attention and spatial attention mechanisms. In the encoding stage of MSR-Net, a multi-scale residual block is introduced to enhance the ability of channel information extraction across varying scales. In the decoding phase, we construct two paths: the introduction of channel attention allows for obtaining the nuanced weighting of the response degree of each channel to key information, and the spatial attention is utilized to extract the global dependence of features to ease the interference of complex background on segmentation performance. When tested on the constructed original pituitary adenoma database, the specificity, IoU, Mcc, and Dice of MSR-Net reached 99.74%, 80.87%, 89.12%, and 89.34%, which were 0.16%, 5.96%, 3.76%, and 3.76% higher than traditional U-Net. Furthermore, we embarked extensive ablative studies on the original dataset to dissect and evaluate the efficacy of each key modules integrated into our architectural framework. Finally, we expanded upon the original dataset to create an enhanced database, and the objective evaluation index further verifies the superiority and advanced nature of MSR-Net. Compared with other most advanced segmentation approaches, MSR-Net shows superior segmentation effect and robustness, as well as great potential in clinical application, which provides an important reference for the future research and development of medical image segmentation.