Weakly Supervised and Embedded Semantic Segmentation for Computer Aided Diagnostics

Abstract

Deep Neural Networks (DNNs) have achieved an exceptionally high level of accuracy in many tasks compared to the previous state-of-the-art (SOTA), thus enabling DNNs to be used for vision-based computer-aided diagnostics (CAD). They offer a great opportunity for physicians and patients. However, to ensure the real-world applicability of visionbased CAD systems, we need to address, on the one side, their dependence on huge pixel-wise annotated datasets, and, on the other side, the reduction of their dependence on expensive and energy-hungry hardware. The research community has published multiple works on weakly supervised semantic segmentation (WSSS), which allows the DNN to perform pixel-wise predictions while being trained on classification labels. However, their downside is reduced prediction quality, further emphasized by the fact that most WSSS research does not target medical applications. Nevertheless, the cost of deployment is another crucial factor in the real-world application of vision-based CAD systems. In computer vision, SOTA typically relies on powerful GPUs for deployment, whereas in the medical sector, the focus is on cost and ease of deployment. Therefore, developing efficient lightweight neural networks to meet the strict hardware limitations of target platforms is unavoidable. Although some work has already been dedicated to lightweight deployment, more efficient DNNs compressed enough to run on embedded platforms at the highest possible prediction quality are always welcome, even more so after the emergence of the transformer architecture.In this regard, this thesis tackles the above-discussed challenges by addressing the shortcomings of WSSS in the medical field and proposing model compression techniques targeted to achieve higher-quality prediction on systems that can run on embedded platforms. More specifically, this research improves the prediction quality of WSSS networks by proposing modules for improved boundary detection and various ensemble approaches. Moreover, this research proposes a comprehensive list of compression approaches to enable the embedded deployment of variations of the popular U-Net and U-Net-shaped transformers.