Weakly Supervised Semantic Cell Segmentation Using Knowledge Distillation

Ivan Vykopal and Ivana Haberova

Supervisor(s): Dr. Lukáš Hudec

Slovak Technical University

Abstract: This study proposes a new approach for semantic cell segmentation that combines the use of neural networks and involving humans in the loop with the aim of improving the current state of digital pathology. The goal is to obtain cell segmentation and classification from heart biopsy images based on inaccurate data and simultaneously to reduce the demands on domain experts - doctors. In the first step, the approach utilizes a segmentation model and a combination of different datasets to detect the nuclei of cells in the patches of whole slide images, which are used to increase the amount of data. The proposed approach employs knowledge distillation, a technique that involves training a smaller "student" model to mimic the output of a larger "teacher" model and their chaining. This is done to overcome the limitations of having a small amount of accurate data and a high proportion of inaccurate annotations and to remove inaccuracies through chaining. The proposed approach is evaluated against traditional methods and shows that it achieves improved performance in terms of semantic cell segmentation. This demonstrates the potential for the approach to be applied in biomedical image analysis, where accurate and precise segmentation is essential for downstream analysis.
Keywords: Computer Vision, Image Processing
Full text:
Year: 2023