Weakly Supervised Segmentation in Medical Imaging: A Counterfactual Approach
Name
Dmytro Shvetsov
Abstract
Deep learning is transforming medical imaging and radiology, notably improving pathology detection in CT and X-ray images. However, the effectiveness of deep learning models in segmentation tasks is often complicated by the requirement for large, annotated datasets. Addressing this challenge, this thesis explores Weakly Supervised Semantic Segmentation enabled by Explainable AI techniques, particularly through generating counterfactual explanations. Our innovative approach, termed Counterfactual Inpainting (COIN), flips classification predictions from abnormal to normal given the input image by erasing identified pathology regions in medical images. For example, when a classifier labels an image as abnormal, COIN produces minimally modified counterfactual example, effectively inverting the original prediction. Our method allows for accurate pathology segmentation without relying on detailed pre-existing masks, using instead much simpler to obtain image-level labels. We validate the effectiveness of COIN by segmenting both synthetic targets and real kidney tumors in TotalSegmentator and Tartu University Hospital CT scan datasets. Our results show that COIN significantly outperforms traditional methods like ScoreCAM, LayerCAM, and RISE, as well as the baseline counterfactual approach by Singla et al. These findings confirm that COIN is a promising method for kidney tumor segmentation in CT images, making a significant advancement in applying deep learning in healthcare domain that is limited by the scarcity of annotated data. The developed approach not only strives for high accuracy required in the medical domain but also reduces reliance on extensive annotated datasets by leveraging semi-automated labeling techniques.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Joonas Ariva, Marharyta Domnich, Dmytro Fishman
Defence year
2024