[1M3-CC] Machine Learning Application to Medical Engineering

[1M3-CC-04] Improving Individually Selectness for Immune Cells using GradCAM

Shoya Kusunose¹, Yuki Shinomiya¹, Takashi Ushiwaka², Nagamasa Maeda³, Yukinobu Hoshino¹ (1. Kochi University of Technology, 2. Kagoshima Uniciersity, 3. Kochi Medical School)

This paper focuses on the analysis of behavior of immune cells for supporting diagnosis. The previous work has proposed to detect immune cells from medical images by recognition frequency space. However, the work has a problem that selects a region including multiple cells as a single cell. This paper aims to relax the problem by focusing the locality of cell shapes. Our proposal uses gradient-weighted class activation mapping (GradCAM) to capture locality characteristics of immune cells. The results shows that the densely inhabited immune cells are correctly selected. Our proposal has some issues that a cell is occasionally detected as multiple ones. We consider that this issue is can be solved by image processing techniques such as non-maximum suppression.