Nuclei Instance Segmentation and Classification for Histopathology Images
Bachelor’s Project-I (Collaboration with TATA Cancer Research Hospital) at the Medical Deep Learning and Artificial Intelligence Lab, IIT Bombay, Guide: Prof. Amit Sethi
Summary: Segmentation and classification of cell nuclei in biopsies helps with the diagnosis of several conditions, saving clinicians’ time. Staining variability of dyes used, severe class imbalance and variation of class labels in open source datasets were some challenges mitigated by us in this project, to consolidate datasets and detect eosinophilia.
Paper published in the 11th International Conference of Bioimaging (BIOSTEC), 2024
Poster published in the Gastroenterology Journal and presented at the Digestive Disease Week 2023, Chicago
- Designed a novel loss function to consolidate fine-grained and hierarchical class labels of PanNuke, MoNuSAC, CoNSeP, achieving improvements on test sets and domain generalization on unseen segmentation and classification datasets
- Combined focal loss and colorjitter augmentation with UNet-based Stardist to detect eosinophilia with 85% accuracy