Related Experiment Video
Updated: Jul 5, 2025
22:49
Optimized Staining and Proliferation Modeling Methods for Cell Division Monitoring using Cell Tracking Dyes
Published on: December 13, 2012
35.7K
Uninformed Teacher-Student for hard-samples distillation in weakly supervised mitosis localization
Claudio Fernandez-Martín1, Julio Silva-Rodriguez2, Umay Kiraz3
1Instituto Universitario de Investigación en Tecnología Centrada en el Ser Humano, HUMAN-tech, Universitat Politècnica de València, Valencia, Spain.
Summary
This study introduces a new weakly supervised method for detecting mitosis in breast cancer histology images using only image-level labels. The approach improves mitosis detection and counting, offering a potential objective tool for assessing tumor proliferation.
Area of Science:
- Computational pathology
- Medical image analysis
- Machine learning for cancer diagnostics
Background:
- Manual mitosis counting is a crucial but challenging biomarker for cancer diagnosis and prognosis.
- Current deep learning methods for mitosis detection are limited by the need for precise centroid labels and suffer from noise from hard negatives.
- Existing methods often require complex multi-stage algorithms to refine pixel-level labels and reduce false positives.
Purpose of the Study:
- To present a novel weakly supervised approach for mitosis detection using only image-level labels on H&E images.
- To introduce an Uninformed Teacher-Student (UTS) pipeline for detecting and distilling hard samples by comparing localizations and annotated centroids.
- To develop an automatic proliferation score mimicking the pathologist-annotated mitotic activity index (MAI).
Main Methods:
- A weakly supervised deep learning framework utilizing image-level labels for mitosis detection.
- An Uninformed Teacher-Student (UTS) pipeline employing strong augmentations to enhance uncertainty and distill hard samples.
- Evaluation on multiple public datasets for mitosis detection and mitotic activity counting in breast histology.
Main Results:
- The proposed framework achieves competitive performance without explicit mitosis location information during training.
- The UTS pipeline demonstrated dataset-specific improvements, enhancing mitosis localization by up to ~4% in cases with less refined annotations.
- The automatic proliferation score showed moderate positive correlation with pathologist-annotated MAI on external datasets.
Conclusions:
- The weakly supervised Uninformed Teacher-Student pipeline effectively leverages strong augmentations to distill uncertain samples.
- The approach demonstrates the feasibility of using image-level labels for mitosis detection and counting.
- The method shows potential as an objective tool for evaluating tumor proliferation in digital pathology.
