I developed a deep learning-based model for automatic segmentation of the spinal cord and intramedullary lesions in patients with traumatic and non-traumatic SCI using T2-weighted MRI scans. The model was developed and evaluated on a diverse, multisite real-world dataset and made publicly available as part of the open-source library for spinal cord MRI data analysis called Spinal Cord Toolbox. Notably, the model represents the first publicly available (open-source) tool for intramedullary lesion segmentation in spinal cord injury patients. As it segments jointly both the spinal cord and the lesions (instead of having two individual models), it goes beyond the current state of the art. The accompanying publication was published as an open-access peer-reviewed paper in the top-tier journal Radiology: Artificial Intelligence (Impact Factor: 8.1 Rank—Radiology, Nuclear Medicine & Medical Imaging: 9/204) and presented at an international conference in Singapore. Then, I updated the model (retrained it with additional new data we gained) and presented it at a workshop focused on applications of medical AI and published it as conference proceedings. The tool gained traction from researchers and clinicians around the world across disciplines. As the model can be applied to both traumatic and non-traumatic SCI patients, it goes beyond the original proposal (initially meant to be designed solely for traumatic SCI) and beyond the current state of the art.
Then, I employed the segmentation tool within a comprehensive analysis pipeline to compute quantitative MRI-derived metrics of the traumatic SCI severity on a large multi-site dataset. As the quantitative MRI-derived metrics might be impacted by MRI image parameters and biological factors (e.g. sex and age), I also focused on developing tools to mitigate these potential biases. Namely, I built a normative database of healthy participants and developed a normalization method allowing to normalize patients’ images relative to this database to account for biological factors such as sex and age and MRI scanner variability. Additionally, we developed an automatic method for quantitative metrics self-normalization to account for differences along the spinal cord anatomy. Finally, I automated the computation of several quantitative metrics, which have been so far computed only manually (which is a subjective, time-consuming task prone to intra- and inter-expert variability), thus going beyond the current state of the art.
After obtaining quantitative MRI-derived metrics using the developed tools, I used them, along with clinical scores and demographic data, to train a machine learning model for predicting patient outcomes. Several clinical outcomes were used as prediction targets, and a range of machine learning algorithms, optimization strategies, feature reduction techniques, and different feature sets were tested to evaluate their ability to predict patients’ clinical outcomes.
