TLDR: Researchers have developed a two-stage AI framework for highly accurate mitosis detection in whole-slide images. The system first uses an improved YOLO11x model to broadly identify potential mitosis candidates, followed by a ConvNeXt-Tiny classifier to precisely filter out false positives. This method achieved an F1-score of 0.882, significantly improving precision and reducing false positives by 54% compared to single-stage baselines, making it more reliable for tumor grading in complex tissue samples.
Accurate mitosis detection is a cornerstone of tumor grading in clinical pathology, yet it remains a challenging task. Manual annotation of mitotic figures in whole-slide images (WSIs) is not only time-consuming but also prone to variability among pathologists due to the subtle and often ambiguous morphology of these cellular divisions. Addressing these critical issues, a new two-stage framework has been developed to significantly enhance the precision and recall of mitosis detection.
Understanding Mitosis Detection
Mitotic figures are crucial indicators for assessing tumor aggressiveness. However, their identification is complicated by several factors. Whole-slide images are enormous, making it difficult to spot small mitotic figures (typically 10-30 pixels) in low-resolution patches. Furthermore, necrotic debris and inflamed cells often mimic the appearance of true mitoses, leading to a high rate of false positives. The diversity of tumor types and specimens across different datasets also poses a challenge to the generalization of detection models.
A Novel Two-Stage Approach
To overcome these hurdles, researchers Jie Xiao, Mengye Lyu, and Shaojun Liu from Shenzhen Technology University have proposed an innovative two-stage framework. This approach combines the efficiency of a detection network with the accuracy of a classification network. The first stage employs an improved YOLO11x model to generate a broad set of potential mitosis candidates, prioritizing recall (ensuring as many true mitoses as possible are captured). The second stage then uses a ConvNeXt-Tiny classifier to meticulously filter out false positives, thereby boosting precision.
How the System Works
The process begins by dividing large whole-slide images into smaller, manageable 512×512 pixel patches with a 20% overlap to ensure no mitotic figures are missed at the edges. These patches are then fed into the improved YOLO11x proposal network. This network has been enhanced with C3k2_LSConv blocks, which use large-kernel convolutions to capture extensive contextual information, and a C2PSA_EMA module with EMA attention, designed to refine feature representation and suppress background noise, especially in complex regions like necrotic zones. By setting a low confidence threshold of 0.2, this stage aims to identify nearly all potential mitotic figures, even if it means including some false positives.
The candidates generated by the YOLO11x stage are then resized to 64×64 pixels and passed to the ConvNeXt-Tiny classifier. This classifier’s primary role is to distinguish true mitotic figures from impostors, effectively reducing the number of false positives. The training of this classifier utilizes a hybrid loss function, combining focal loss to handle class imbalance and contrastive loss to enhance the separation of features between true mitoses and hard negative samples.
Impressive Performance Gains
Evaluated on a combined dataset including MIDOG++, MITOS_WSI_CCMCT, and MITOS_WSI_CMC, the proposed two-stage framework achieved an F1-score of 0.882. This represents a significant improvement of 0.035 over the single-stage improved YOLO11x baseline and 0.055 over the basic YOLO11x. The most notable gain was in precision, which jumped from 0.762 to 0.839, while maintaining a comparable recall. This translates to a remarkable 54% reduction in false positives compared to the basic YOLO11x, making the detection results much more reliable for clinical applications.
The researchers highlight that this synergy between YOLO11x’s efficiency in candidate generation and ConvNeXt’s robustness in classification directly addresses the unique challenges of the MIDOG 2025 Track 1, particularly in complex and heterogeneous tissue contexts.
Also Read:
- SDF-YOLO: A Focused AI Approach for Robust Mitotic Figure Detection in Pathology
- Adaptive AI Strategies Advance Atypical Mitosis Classification in Pathology
Looking Ahead
Future work will explore dynamic threshold adjustment mechanisms that can adapt to specific WSI features, such as the proportion of necrotic regions or tumor type. Additionally, integrating domain adaptation modules is planned to further improve the framework’s generalization across diverse clinical scenarios, ensuring its applicability in a wider range of diagnostic settings.
For more in-depth information, you can read the full research paper here.
