Mitosis Detection, Fast and Slow (MDFS)
About
- Histopathology region of interest cropout (ROI cropouts of histopathology images)
- Mitotic Figures (Detected mitotic figures, where each point represents the center of a mitotic cell. Additionally or optionally, confidence scores can be returned.)
Challenge Performance
| Date | Challenge | Phase | Rank |
|---|---|---|---|
| Aug. 22, 2022 | MIDOG2022 | Preliminary Test Phase (OVER) | 1 |
| Sept. 5, 2023 | MIDOG2022 | Final Test Phase (Task 1, Without additional data) | 1 |
Model Facts
Summary
We suggest an effective and reliable dual-phase system for the detection of mitosis, which involves an initial identification of potential mitotic figures followed by the fine-tuning of these candidates. The proposed model for the preliminary segmentation of candidates, named EUNet, is characterized by rapid performance and accuracy, owing to its structural design. EUNet is capable of delineating candidate areas at a reduced resolution, which significantly accelerates the detection process. Subsequent refinement of the initially identified candidates is carried out using a more complex classifier network known as EfficientNet-B7 during the second phase.
This algorithm was developed for MIDOG22 challenge.
Mechanism
-
Target Population: The algorithm we developed is primarily intended for use by pathologists, computational pathologists, and researchers. Its function is to aid in the analysis and interpretation of histopathological data, specifically, it targets the process of identifying mitotic cells within histology samples.
-
Algorithm Description: Our algorithm operates through a two-phase process for detecting mitosis in cellular structures. In the first stage, the EUNet model is utilized for the preliminary identification of possible mitotic candidates. This model excels in speed and precision due to its architecture, enabling it to segment potential mitotic figures even at lower resolutions, therefore enhancing the speed of the detection process. The second stage of the algorithm employs a sophisticated classifier network, EfficientNet-B7, to refine the segmented candidates. It cross-checks and confirms the initial identifications, improving the accuracy of the final results. Additionally, to prevent variations caused by domain shifts, the algorithm incorporates domain-neutral techniques ensuring its robustness.
-
Inputs and Outputs: The input to our algorithm consists of standard histology images prepared with Hematoxylin and Eosin (H&E) staining, a common staining method in medical histology. As for the output, the algorithm provides a list of point detections each associated with a probability score indicating the likelihood of it being a mitotic figure. This way, the pathologists, computational pathologists, and researchers can better focus their analyses, speeding up and increasing the accuracy of their work.
Validation and Performance
Results for this algorithm on MIDOG21 and MIDOG21 test sets can be found on their leaderboards. https://midog2022.grand-challenge.org/evaluation/final-test-phase-task-1-without-additional-data/leaderboard/
Uses and Directions
This algorithm was developed for research purposes only. This algorithm does not work on IHC images properly.
Warnings
Common Error Messages
Information on this algorithm has been provided by the Algorithm Editors, following the Model Facts labels guidelines from Sendak, M.P., Gao, M., Brajer, N. et al. Presenting machine learning model information to clinical end users with model facts labels. npj Digit. Med. 3, 41 (2020). 10.1038/s41746-020-0253-3