Using graph cuts for anatomical priors to segment Pulmonary Embolus : label generation for deep model training

Details

Supervisors

Macq, Benoît ; De Vleeschouwer, Christophe

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en mathématiques appliquées, à finalité spécialisée

Abstract

This thesis presents a novel Pulmonary Embolus (PE) detection and segmentation approach by combining classical computer vision and deep learning techniques. Pulmonary Embolism is a life-threatening medical condition characterized by the obstruction of pulmonary arteries, often tricky and time-consuming to diagnose accurately. The study addresses the possibility of generating segmentation of PE candidates as labels for a weakly supervised model. We also explore the potential enhancement in performance by incorporating prior anatomical segmentation in the chest. Implementing a classical algorithm for PE segmentation, our methodology involves generating superpixels and employing graph cuts for prior anatomical segmentation, all within the broader context of generating labels for training a weakly supervised learning framework. Our findings suggest that the presented methodology, while promising, still needs refinement to reduce the number of false positives and improve the segmentation of PE regions. This work forms a foundation for further research to improve PE diagnosis and treatment by enhancing the efficacy of deep learning models without the need for extensively human-annotated CT scan datasets.