Semantic Segmentation for Improved Cell Nuclei Analysis
Joren Regan, C Ramesh, Saurabh Gupta, Ankur Sharma, Duc-Son Pham, Aneesh Krishna · 2023
Generalised cell nuclei segmentation on Hematoxylin and Eosin stained tissue allows for faster analysis of tissue slides by pathologists. Improving the quality of automated processes in turn allows for more accurate diagnosis and prognosis. Existing methods make use of synthetic or private datasets that do not allow for collaboration by other researchers. Evaluation metrics vary among papers, making model performance hard to gauge. To incite further research in the space, this study proposes evaluating existing preprocessing techniques and neural networks on the publicly available datasets: MoNuSeg, MoNuSAC, TNBC and CryoNuSeg. The U-Net architecture is compared as the baseline against DeepLabV3+, YOLOv8 and STDC. A qualitative analysis is performed on predictions to assess the shortcomings of existing methods. Evaluation metrics are discussed to reduce ambiguity when training, and more robust metrics are outlined. This study contributes multiple resources to assist further research utilising public data, including end-to-end dataset preparation, along with training and evaluation pipelines for the aforementioned datasets and models.