Injecting Prior Knowledge in Medical Image Interpretation

Computer-Aided Diagnosis (CAD) systems are vital in medical image analysis, assisting clinicians in interpreting complex imaging data to diagnose conditions such as cancer. However, while traditional CAD systems excel at detecting these patterns, they often struggle to follow a line of reasoning needed to draw inferences. This gap has led to growing interest in neurosymbolic AI. By integrating neural and symbolic approaches, neurosymbolic AI—especially through frameworks like Logic Tensor Networks (LTNs)—has the potential to enhance CAD systems by making them more robust and interpretable. This advancement addresses key issues of reliability and transparency, which are crucial in high-stakes fields like healthcare. For instance, it can help us to better address multi-class classification challenges in mammography image analysis where the primary objective is to enhance the automated detection and classification of various mammographic findings, including masses and calcifications, which are critical for early breast cancer diagnosis. A rule can be established stating that a mass with a spiculated margin strongly indicates a high risk of malignancy. A knowledge base K is then formed by logical rules derived from domain-specific data, which are used to define an optimization problem. This is achieved by applying a formula aggregation operator to compute the overall truth value of K. Several configurations of operators have been proposed by researchers, balancing effectiveness, numerical stability, and adaptability to various formulas. The principle aim of this thesis is to select the proper configuration of fuzzy operators for grounding formulas in the logarithmic space and possibly introduce a finer resolution. However, the presence of imbalanced data can still impact performance, as it may lead the model to favor predictions for the majority class while overlooking the minority class. Hence, as the second objective, additional analyses are conducted to address the issues caused by data imbalance. To mitigate this, focal loss is applied as a specialized loss function, designed to down-weight the loss for well-classified examples, enabling the model to focus more on harder, minority-class examples. This approach rebalances the learning process by amplifying the importance of underrepresented classes, helping to enhance the model’s overall reliability and robustness across all classes.