towards a trustworthy data-driven clinical decision support system: breast cancer use-case
abstract
artificial intelligence (ai) research has emerged as a powerful tool for health-related applications. with the increasing shortage of radiologists and oncologists around the world,
developing an end-to-end ai-based clinical decision support (cds) system for fatal disease
diagnosis and survivability prediction can have a significant impact on healthcare professionals as well as patients. such a system uses machine learning algorithms to analyze
medical images and clinical data to detect cancer, estimate its survivability and aid in
treatment planning. we can break the cds system down into three main components:
the computer-aided diagnosis (cad), the computer-aided prognosis subsystem (cap)
and the computer-aided treatment planning (catp). the lack of trustworthiness of these
subsystems is still considered a challenge that needs to be addressed in order to increase
their adoption and usefulness in real-world applications. in this thesis, using the breast
cancer use case, we propose new methods and frameworks to address existing challenges
and research gaps in different components of the system to pave the way toward its usage
in clinical practice.
in cancer cad systems, the first and most important step is to analyze medical images to identify potential tumors in a specific organ. in dense prediction problems like
mass segmentation, preserving the input image resolution plays a crucial role in achieving good performance. however, this resolution is often reduced in current convolution
neural networks (cnn) that are commonly repurposed for this task. in chapter 3, we
propose a double-dilated convolution module in order to preserve spatial resolution while
having a large receptive field. the proposed module is applied to the tumor segmentation
task in breast cancer mammograms as a proof-of-concept. to address the pixel-level class
imbalance problem in mammogram screenings, different loss functions (i.e., binary crossentropy, weighted cross-entropy, dice loss, and tversky loss) are evaluated. we address
the lack of transparency in current medical image segmentation models by employing and
quantitatively evaluating different explainability methods (i.e., grad-cam, occlusion sensitivity, and activation visualization) for the image segmentation task. our experimental
analysis shows the effectiveness of the proposed model in increasing the similarity score and
decreasing the miss-detection rate. [...]