Deep Learning Methods for Leveraging Data Scarcity and Enforcing Prior in Healthcare Image Analysis

Deep learning has been transforming several fields over the past decade with its ability to learn complex data representations and solve difficult tasks. In particular, it has shown significant potential in healthcare image analysis, automating diag- nostic and decision-making processes. Classification and segmentation tasks have seen substantial advancements, enabling faster and more accurate results. However, healthcare data often come with unique hurdles, such as high annotation costs, ethi- cal constraints, and limited availability of datasets for rare medical conditions. Fur- thermore, healthcare images are produced from various acquisition modalities, with high variability among them. Besides, neural networks are often regarded as ”black boxes”, for the lack of transparency and difficulty of explaining their predictions, slowing down their adoption in the healthcare domain. These challenges necessitate innovative solutions to ensure the robustness, reliability, and interpretability of deep learning models. In this thesis, we address two fundamental challenges in applying deep learning to healthcare image analysis: 1. Learning with Data Scarcity: Implementing training strategies with limited data constraints, leveraging methods like transfer learning and self-supervised learning, to overcome Data Scarcity in applications including food image clas- sification, prostate magnetic resonance images segmentation, and cell filopodia segmentation. 2. Enforcing Medical Priors: Incorporating domain-specific knowledge into clas- sification pipelines and aligning the prediction with doctors’ reasoning to im- prove accuracy and add interpretability to the obtained results. This approach addresses key challenges in tasks such as autoimmune bullous disease classifi- cation, midpalatal suture maturity assessment, and lung adenocarcinoma tumor pattern identification.