Ph.D. Candidate: Görkem Polat
Program: Medical Informatics
Date: 17.07.2023 / 12:30
Place: A-212
Abstract: Ulcerative colitis (UC) is a chronic inflammatory bowel disease that presents significant diagnostic and management challenges for clinicians. Accurate assessment of disease severity is crucial for guiding appropriate treatment strategies and improving patient outcomes. The Mayo endoscopic score (MES) is a widely used tool for evaluating UC severity; however, the assessment process relies heavily on subjective interpretation, leading to substantial intra- and inter-observer variability.
In this thesis, we present a novel loss function, termed Class Distance Weighted Cross Entropy (CDW-CE) loss, for the automated assessment of UC severity, harnessing the power of convolutional neural networks (CNN) to analyze endoscopic images of the colon. CDW-CE addresses the limitations of conventional cross-entropy loss functions in ordinal classification problems.
The proposed CDW-CE loss effectively penalizes mispredictions based on their distance from the true class, taking into account the inherent ordinal relationships among the output classes. CDW-CE has been evaluated against other loss functions and consistently outperformed them across various performance metrics and CNN architectures. Moreover, the proposed approach enables the generation of more accurate class activation maps, which can be utilized to explain model predictions —an essential aspect of translating these techniques into clinical practice. To demonstrate the generalizability of the proposed approach, it is also tested on a diabetic retinopathy dataset and got similar results, indicating that the proposed approach can be used in other applications presenting ordinal classes. The dataset created for this study, named Labeled Images for Ulcerative Colitis, is the largest publicly available labeled UC dataset to date.