HealNNet-Lesions: A Deep Learning Framework for Spinal Lesion Detection Demonstrating X-Ray Images as a Screening Alternative to MRI and CT with Deep Learning in Radiology

Radiographs are the most common imaging tool for identifying spinal lesions. The detection of these lesions can be challenging for radiologists. This work details the development and evaluation of a deep learning framework, named HealNNet-Lesions, for the detection of spinal lesions. It also explores the use of artificial intelligence and deep learning to potentially serve as an alternative for diagnosis and initial screening of specific findings in radiology instead of expensive and less accessible Magnetic Resonance Imaging (MRI) and Computed Tomography (CT) scans with more affordable and accessible X-ray scans, particularly in rural and sub-urban India and other low-middle income countries (LMICs). A training dataset was constructed using 4502 Spine X-ray images from the VinDr-SpineXR Dataset. This dataset included findings in eight categories, six of which were unique spinal lesions, along with "Other Lesions" and "No Finding" categories. Using this dataset, a deep learning classifier was trained to determine if a spine scan was abnormal or normal, and a detector, named HealNNet-Lesions, was also trained to localize the lesions. The classifier was evaluated on a test dataset of 468 X-ray scans and achieved an area under the receiver operating characteristic curve (AUROC) of 88.84% and an accuracy of 71.37% for the image classification task. For lesion detection, HealNNet-Lesion achieved a mean average precision (mAP@0.5) of 43.73%. Since spinal lesions often require MRI and CT scans for their diagnosis, this framework demonstrates how deep learning can reduce the need for these expensive and less accessible imaging methods by supporting clinical decision-making where MRI/CT are unavailable with an X-Ray. These results establish a benchmark and serve as a proof of concept to encourage future research in this area.