Deep Learning-Based Hippocampal Segmentation and MTA Classification Using U-Net with ResNet-50 Backbone

Medial Temporal Atrophy (MTA) is a key biomarker in the early diagnosis of dementia. However, its assessment through manual inspection of MRI scans is subjective, time-consuming, and prone to inter-observer variability. This creates the need for automated systems that can provide accurate, consistent, and clinically interpretable evaluations. This study aims to develop a hybrid deep learning framework that integrates U-Net with a ResNet-50 backbone for simultaneous hippocampal segmentation and MTA grading, thereby reducing diagnostic subjectivity and bridging the gap between image processing and clinical interpretation. The main contribution of this work lies in the dual functionality of the proposed architecture: not only producing precise segmentation masks of the hippocampal region but also classifying the degree of atrophy into MTA scores (0–4), which previous studies on hippocampal segmentation have not addressed. The proposed method employs a U-Net for pixel-level segmentation, enhanced with a ResNet-50 backbone to stabilize gradient propagation and enrich feature representation during encoding. Results demonstrated excellent performance, achieving a training accuracy of 99.9% with strong convergence between training and validation curves. On a test set of 32 coronal MRI slices, the model correctly classified 26 samples, misclassifying only 6. Overall, the proposed U-Net with ResNet-50 backbone provides an accurate and reliable end-to-end solution for hippocampal segmentation and MTA grading. Its clinical performance demonstrates parity with expert radiologists, underscoring its potential as a decision-support tool in dementia diagnosis. Future work will focus on extending this framework to 3D U-Net architectures, enabling the integration of volumetric MRI features to enhance robustness and generalizability across diverse datasets further