Deep Transfer Learning for Automated Detection of Spinal Lesions from CT scans

In this thesis we implement an automated Computer-Aided Detection (CADe) system for spine lesions using Computed Tomographies (CTs) and Convolutional Neural Networks (CNNs). To this end, we conceptualize an algorithmic approach for the whole process of extraction and processing of the vertebrae from CT scans, which also manages the detection step for the whole vertebral body. For training and testing purposes, we generated a dataset composed of several CTs in collaboration with the Rizzoli Orthopaedic Insitute of Bologna, Italy. The vertebrae, either healthy or containing lesions (e.g. metastases, primary tumors, lytic and sclerotic lesions) were extracted from CT scans with a toolbox developed ad hoc to automatize the process. The resulting dataset is composed of slices from the previously extracted volumes containing the vertebrae. Slices were processed with contrast enhancement and data augmentation techniques, and subsequently used to train the Neural Network. For the purpose of detection, we perform an in-depth comparative study by implementing 4 pre-trained networks and exploiting Transfer Learning techniques. To prove the great advantages of Transfer Learning, we show how the pre-trained networks outperform a network trained from scratch, reaching 95.97% accuracy and F1 score of 94.22%. Finally, we equip the CADe system with an intuitive Graphical User Interface (GUI) to allow physicians to use the automated detection software as a support tool for diagnoses on new patients.