Enhancing Mitotic Cell Segmentation: A Transformer Based U-Net Approach

Mitosis segmentation plays a vital role in early cancer detection, facilitating the accurate identification of dividing cells in histopathology images. Manual mitosis counting is time-consuming and subjective, prompting the need for automated approaches to improve efficiency and accuracy. In this study, we have developed a transformer-based U-Net model that combines the effectiveness of transformers which were originally designed for natural language processing (NLP) tasks, with the efficiency of the U-Net architecture to effectively capture both high-level and low-level features in histopathology images. We train and evaluate the model on the GZMH dataset and compare its performance against other deep models such as U-Net, U-Net++ and Mobilenetv2-based U-Net. The results demonstrate that transformer-based U-Net model is better in terms of accuracy, recall, precision, F1-score and Dice coefficient. This study represents a significant advancement in mitosis segmentation, contributing to improved cancer detection and prognosis.