Mitosis Detection for Breast Cancer Pathology Images Using UV-Net

TL;DR

This paper presents UV-Net, a deep learning architecture optimized for high-resolution mitosis detection in breast cancer pathology images, addressing challenges of data diversity and object similarity.

Contribution

The paper introduces UV-Net, a novel architecture with feature preservation and stain normalization, improving mitosis detection in complex pathology images.

Findings

Achieved an F1 score of 0.6721.

Enhanced detection accuracy with high-resolution focus.

Demonstrated robustness across diverse datasets.

Abstract

The difficulty of detecting mitosis and its similarity to non-mitosis objects has remained a challenge in computational pathology. The lack of publicly available data has added more complexity. Deep learning algorithms have shown potentials in mitosis detection tasks. However, they face challenges when applied to pathology images with dense medium and diverse dataset. This paper introduces an optimized UV-Net architecture, developed to focus on mitosis details with high-resolution through feature preservation. Stain normalization methods are used to generalize the trained network. An F1 score of 0.6721 is achieved using this network.