UTILIZING TRANSFER LEARNING FOR VERTEBRAL FRACTURES DETECTION

Abstract

Spinal fractures are a prevalent type of fracture that have led to serious health issues which include difficulty in movement and permanent pain. Vertebral fractures are frequently present in most CT scans taken for abdominal health issues, and so they have rarely been diagnosed. Moreover, manual detection in medical images is time-consuming and requires specialized training. Thus, aiming for early and automated vertebral fracture detection is crucial for effective and fast treatment. Machine learning automated techniques can be utilized for fracture detection while deep learning models have proven their power in diagnosing different types of diseases. Specifically, transfer learning models have proven their effectiveness in diseases’ detection from limited medical data benchmarks. For this reason, the project suggests the use of transfer learning models to effectively diagnose vertebral fractures from a CT scan dataset at AUBMC. Five different deep architectures models (ResNet26, ResNet-RS-50, Inception_ResNet_v2, Swin_S3_Tiny, and ConvNeXT_Tiny_in22k) have been selected for investigation to diagnose fracture presence from the AUBMC CT scan dataset after passing through a series of pre-processing. Segmenting the vertebra along with classical augmentation were two important pre-processing steps that have improved the binary classification performance in all models. The results have shown that ConvNeXT_Tiny_in22k model has attained the highest testing accuracy of 81% without segmentation and 84.8% with segmentation upon selecting the same number of CT scan images. The dataset was then extended and segmented where the ConvNeXT model outperformed with a testing accuracy of 96.4%.