Application of Machine Learning for the Detection and Classification of Spinal Pathologies Using YOLO12x

Abstract

The integration of Machine Learning (ML) in healthcare has driven significant advancements in medical imaging analysis and automated diagnosis. One of the ongoing challenges in this area is the detection and classification of multiple abnormalities in a single radiographic examination. This study focuses on evaluating the effectiveness of the YOLO12x (You Only Look Once, version 12x) model in detecting and classifying seven distinct spinal pathologies: disc space narrowing, foraminal stenosis, osteophytes, other lesions, spondylolisthesis, surgical implant, and vertebral collapse. Trained and evaluated on 5,006 radiographs from the VinDrSpineXR dataset, the model without preprocessing achieved average values of precision 0.9677, recall 0.8671, F1-score 0.9127, mAP@0.5 0.9271, and mAP@0.5:0.95 0.8501. Class-wise, the strongest performance was observed in spondylolisthesis, surgical implant, and vertebral collapse <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\mathbf{F} 1 \geq 0.94; \mathbf{m A P} {@} 0.5 \geq \mathbf{0. 9 4})$</tex>. The normalized confusion matrix indicated high accuracy for foraminal stenosis <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(90 \%)$</tex>, osteophytes <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(89 \%)$</tex>, surgical implant (85%), and vertebral collapse (82%), with lower accuracy for other lesions (72%). A preprocessing pipeline based on CLAHE, Sobel filtering, and intensity normalization did not yield improvements (precision 0.9339; recall 0.7747; F1 0.8454; mAP@0.5 0.8722; mAP@0.5:0.95 0.6917), although it slightly increased precision for foraminal stenosis. These results indicate that YOLO12x reliably detects multiple spinal pathologies in radiographs without preprocessing, supporting real-time clinical workflows; future work should explore class balancing and data augmentation techniques to improve recall in more challenging categories.

Ähnliche Arbeiten