Deep Learning-Based Segmentation and Classification of Histological Colon Cells

Patrik Kozlík

Supervisor(s): prof. Ing. Vanda Benešová, PhD.

Slovak Technical University

Abstract: Deep neural networks have become important in the research of medical applications, particularly in histology. One of the important research areas is the usage of deep neural networks in the diagnostics of diseases such as Crohn's disease and Ulcerative Colitis. In these types of diseases, the correct detection of specific cell types and cell features is crucial. To integrate this domain-specific knowledge directly from pathologists, we propose a customizable system made of three connected modules: segmentation, filtration, and classification. The segmentation module uses the AttentionUNET architecture to find cell boundaries. The filtration module contains a stack of filters suitable to specific cell features, such as color, shape, and area. These filters are used to eliminate irrelevant cells from the predicted segmentation mask. The classification module uses the ResNet34 architecture for multi-class classification tasks. Through experimentation involving custom loss functions and attention modules, we found that the filtration module is well-suited for elimination of irrelevant cells from the predicted mask. The segmentation module achieves a Dice score of 84.18% and an F1-score of 90.08%. However, the classification module exhibits an accuracy of approximately 72%, primarily due to limited annotated data. Nonetheless, our solution proves effective in scenarios with constrained training data, as the filtration module aids with process of prediction by filtering out irrelevant cells.
Keywords: Computer Vision, Human-Computer Interaction, Image Processing
Full text:
Year: 2024