Content area
This thesis investigates machine learning techniques, specifically traditional machine learning and deep learning, to improve diagnostic accuracy, interpretability, and computational efficiency in healthcare applications. It addresses tasks such as multi-modal breast cancer diagnosis, Alzheimer’s Disease (AD) diagnosis, depression detection in AD patients, and functional outcome prediction after endovascular thrombectomy following an acute ischemic stroke. This work employs Generative Adversarial Networks (GANs) to augment digital mammography images, focusing on intra- and cross-modality synthesis. Results indicate that GAN-generated data enhances model performance within a single modality, though challenges remain in preserving diagnostic features during cross-modality synthesis. Additionally, the thesis introduces
Forest: Net, a hybrid model that integrates Decision Tree Ensembles with Artificial Neural Networks to enhance interpretability without sacrificing predictive power. The tree-based structured sparsity of
Forest: Net enables it to focus on clinically relevant features, making it computationally efficient and suitable for resource-limited environments. Evaluated on diverse datasets,
Forest: Net shows promising results on non-visual healthcare data, providing a unified tool for high accuracy and interpretability. While its applicability is limited in complex visual data, it achieves high accuracy on simpler visual data (e.g., MNIST). Moreover, enforcing its structured sparsity during training demonstrates regularizing properties when
Forest: Net is used as the classification head of Convolutional Neural Networks. This thesis establishes a foundation for developing interpretable, efficient, and reliable AI tools in healthcare. Since interpretability fosters clinician trust, this work advances ethical AI practices and responsible AI integration in healthcare.