SoftCTM: Cell detection by soft instance segmentation and consideration of cell-tissue interaction
This work addresses cell detection in computational pathology, which is incremental as it builds on existing methods by optimizing ground truth formats and adding tissue interaction features.
The paper tackled cell detection in histopathology images by exploring ground truth formats and incorporating cell-tissue interactions, resulting in a model that achieved a 0.7172 mean F1-Score on the OCELOT test set, ranking third in the OCELOT 2023 Challenge.
Detecting and classifying cells in histopathology H\&E stained whole-slide images is a core task in computational pathology, as it provides valuable insight into the tumor microenvironment. In this work we investigate the impact of ground truth formats on the models performance. Additionally, cell-tissue interactions are considered by providing tissue segmentation predictions as input to the cell detection model. We find that a "soft", probability-map instance segmentation ground truth leads to best model performance. Combined with cell-tissue interaction and test-time augmentation our Soft Cell-Tissue-Model (SoftCTM) achieves 0.7172 mean F1-Score on the Overlapped Cell On Tissue (OCELOT) test set, achieving the third best overall score in the OCELOT 2023 Challenge. The source code for our approach is made publicly available at https://github.com/lely475/ocelot23algo.