Nuclei Classification of Inflammatory Cells in Pediatrics: A Novel Transfer Learning Approach for Digital Histopathology

The field of Digital Histopathology has seen many significant advancements in the past decade, and this is primarily because of the developments that have been seen in Artificial Intelligence, pertaining to the segmentation and classification of cell types in Whole Slide Images (WSI) of tissues. The focus of these developments has been on neoplastic or cancerous cell types in adults as the primary area of research. In this work, we propose a transfer learning approach to improve the classification of inflammatory cells for pediatric patients, as it is an emerging research area that is gaining momentum in the medical field. We conduct a review to identify the existing models in this field and propose a novel approach to improve the model's performance through the usage of transfer learning. We chose the PanNuke dataset, as it provides labeled tissue samples that have been expertly reviewed and have been shown to encompass real-world variations that would be experienced in a real-world setting, making it a more practical choice for training the model. We showcase that our SqueezeNet model developed on Fast.ai archives state-of-the-art performance on the PanNuke dataset, with an overall weighted average F1-Score of 89.2% on the binary classifier (between inflammatory cells and all other types of cells.) It also achieves 79.3% on the multi-class classifier (between all five cell types included in the PanNuke dataset.) We also show that our score achieves an improvement of 41% in the weighted average F1-Score against the previous state-of-the-art Hover-Net model for the Neoplastic, Epithelial, Inflammatory, and Connective cell types. Furthermore, we demonstrate achieving state-of-the-art sensitivity and precision scores in all trained cell types for the classification task for individual class performance, and the overall model weighted average.