<![CDATA[Manually selecting MRI slices is a significant bottleneck in clinical workflows. This issue is worsened by inconsistent naming conventions and variable acquisition protocols across institutions and radiologists, often leading to redundant efforts and potential oversights during medical image data preprocessing. This study introduces a fully automated, four-level hierarchical classification system specifically designed to intelligently filter and select clinically relevant spinal MRI slices directly from raw DICOM series. Our primary objective is to streamline the initial stages of radiological assessment, ensuring that only pertinent images are presented for subsequent analysis and review. We thoroughly evaluated the performance of modern, efficient deep learning architectures, including EfficientViT, MobileNetV4, and RepViT, benchmarking them against a robust ResNet-18 baseline. The proposed pipeline systematically refines its analysis through a structured hierarchy: it first broadly identifies the anatomical region, then precisely classifies the spine location and specific view (axial, sagittal, or coronal). Subsequently, it categorizes the imaging contrast, and finally, confirms the presence of the spinal cord. Our comprehensive experimental results reveal that the EfficientViT-based model achieved the highest end-to-end F1-score of 0.8357, demonstrating robust accuracy across all classification levels. Furthermore, its average inference speed of 9.17 ms per image highlights its computational efficiency. This automated pipeline offers an effective and computationally efficient solution for speeding up initial medical image preprocessing, ensuring subsequent analytical tasks are performed on accurately selected, clinically relevant data.]]>
Copyrights © 2025
