TLDR: Researchers have developed a novel AI model for the MIDOG 2025 challenge, aiming to improve mitotic figure detection in cancer pathology. Their approach extends the FCOS object detector by integrating a Feedback Attention Ladder CNN (FAL-CNN) and a fusion network. This composite system classifies mitotic figures and refines bounding box predictions to reduce false positives, enhance accuracy, and improve generalizability across diverse datasets. While initial F1 scores were lower than the baseline, the model is expected to perform better on unseen data and offers explainability through its attention mechanism.
Accurate detection of mitotic figures, which are cells undergoing division, is crucial for grading cancer and predicting its progression in histopathological images. However, this task is challenging for Artificial Intelligence (AI) due to variations in tissue appearance, staining, and how mitotic figures look across different datasets. The MIDOG 2025 challenge aims to find solutions that can generalize well across these variations for mitotic segmentation and grading.
Researchers Andrew Broad, Jason Keighley, Lucy Godson, and Alex Wright from the National Pathology Imaging Cooperative, UK, have introduced a new method to improve mitotic figure detection. Their approach, detailed in their paper MIDOG 2025: Mitotic Figure Detection with Attention-Guided False Positive Correction, extends an existing object detection system called FCOS (Fully Convolutional One-Stage Object Detector).
A Novel Approach to Mitosis Detection
The core of their innovation is a composite model that integrates a Feedback Attention Ladder CNN (FAL-CNN) with the FCOS detector. The FAL-CNN is specifically designed to classify whether a detected figure is a normal or abnormal mitotic figure. This classification information then feeds into a separate ‘fusion network’. This fusion network is trained to make precise adjustments to the bounding boxes (the predicted outlines around detected objects) initially suggested by the FCOS model.
The primary goal of this sophisticated network is to significantly reduce the number of ‘false positives’ – instances where the system incorrectly identifies something as a mitotic figure. By doing so, the model aims to enhance the overall accuracy of object detection and, crucially, improve its ability to perform well on new, unseen datasets, making it more generalizable.
How the System Works
The image processing pipeline consists of four main components working in harmony:
- FCOS Object Detector: This is the initial stage, a fully convolutional network that identifies potential mitotic figures and draws bounding boxes around them, along with a confidence score.
- FAL-CNN Classifier: This component takes small sections (mini-patches) centered on the FCOS-predicted bounding boxes. It uses a hierarchical feedback attention mechanism to classify these patches as either mitotic or non-mitotic. It also generates a ‘spatial attention map’ that highlights important cellular features, guiding the system’s focus. The FAL-CNN is built upon the VGG19 architecture and includes a feedback pathway that enhances its response to salient image regions.
- Fusion Network: This is where the magic of refinement happens. It takes the spatial attention map and the probability of mitosis (pmitosis) from the FAL-CNN, along with the original FCOS scores. Using a series of connected layers, it calculates precise adjustments – coordinate offsets and score multipliers – to fine-tune the position and confidence score of the FCOS-predicted bounding boxes. The fusion network is trained while the FCOS and FAL-CNN parts remain fixed.
- Mini-patch Sampling: This step involves extracting 56×56 pixel samples from the input image, centered around the bounding box locations predicted by FCOS. These mini-patches are then fed into the FAL-CNN for classification.
Training and Performance
The FAL-CNN and fusion network were trained using the MIDOG++ dataset, which is publicly available. The FAL-CNN was specifically trained on a balanced set of nearly 12,000 mitotic and 12,000 non-mitotic figure patches. For the MIDOG 2025 Track 1 Evaluation phase, the composite model achieved an F1 score of 0.655. Interestingly, the baseline FCOS model, when trained on the same MIDOG++ data, achieved a higher F1 score of 0.767 in this initial test phase.
Also Read:
- Enhancing Mitosis Detection in Digital Pathology with a Hybrid AI Approach
- SDF-YOLO: A Focused AI Approach for Robust Mitotic Figure Detection in Pathology
Looking Ahead
While the initial performance was lower than the baseline, the researchers are optimistic about their model’s potential. They hypothesize that their attention-guided approach will be more robust and less affected by variations in datasets and different imaging domains. The true test of this generalizability will come with final analysis on unseen test sets. Furthermore, the FAL-CNN’s feedback mechanism offers a valuable benefit: it can highlight the specific features that lead to its classifications, providing a degree of explainability for its decisions, which is a significant advantage in AI applications for pathology.
