TLDR: Researchers developed a robust, real-time mitotic figure detection method using an efficient RTMDet-S model and a “bag of tricks” including multi-domain training, balanced sampling, data augmentation, and hard negative mining. This approach achieved an F1 score of 0.81 on the MIDOG 2025 challenge, outperforming larger models, and offers a practical balance of accuracy and speed for clinical deployment in cancer diagnosis.
Detecting mitotic figures (MF) in histopathology images is a critical step in evaluating various cancer types. However, this task is often challenging due to significant variations in slide scanners, staining protocols, tissue types, and the presence of artifacts. Researchers Christian Marzahl and Brian Napora from Gestalt Diagnostics have introduced a novel approach to tackle these difficulties, presenting a collection of training techniques – aptly named a “bag of tricks” – that enable robust, real-time MF detection across diverse domains.
Their method is built upon the efficient RTMDet single-stage object detector, specifically utilizing the small variant, RTMDet-S. This choice was deliberate, aiming for an optimal balance between high accuracy and rapid inference speed, making the solution highly suitable for practical clinical deployment. The core of their strategy addresses common issues like scanner variability and tumor heterogeneity through extensive training data gathered from multiple domains, a carefully balanced sampling strategy, and targeted data augmentation. Furthermore, they employ a technique called hard negative mining, focusing on necrotic and debris tissue, to significantly reduce false positives – instances where the system incorrectly identifies non-mitotic elements as MFs.
The model’s performance was rigorously evaluated using a grouped 5-fold cross-validation across multiple MF datasets, where it consistently achieved an F1 score ranging between 0.78 and 0.84. In a preliminary test set for the MItosis DOmain Generalization (MIDOG) 2025 challenge, their RTMDet-S based approach achieved an impressive F1 score of 0.81. This result is particularly noteworthy as it surpassed the performance of larger, more complex models and demonstrated remarkable adaptability to new, previously unseen domains. The solution represents a practical trade-off between detection accuracy and processing speed, making it an attractive candidate for real-world clinical adoption.
The researchers emphasize the growing importance of machine learning algorithms in digital pathology for quantifying mitotic activity. They acknowledge that algorithm performance can deteriorate under “domain shifts” caused by different whole-slide scanners, staining protocols, tissue types, or even species. These variations lead to challenges such as color appearance differences, heterogeneous tumor morphology, and distractors like artifacts or inflammation that can be mistaken for mitoses. Previous challenges, including MIDOG 2021 and 2022, have highlighted the severity of this problem.
To build a robust model, the team utilized a comprehensive combination of four MF detection datasets: MIDOG++, Canine Cutaneous Mast Cell Tumor (CCMCT), Canine Mammary Carcinoma (CMC), and TUPAC16. These datasets collectively cover a broad spectrum of domains, encompassing various human and canine tumor types, and images originating from multiple laboratories and scanners, thereby ensuring substantial domain variability. To further refine the model’s discriminative capabilities, they incorporated hard negative mining. This involved using external datasets like HistAI SPIDER, which focuses on necrotic tissue, and NCT-CRC-HE-100K, containing debris patches. These non-MF patches were strategically added as challenging negative examples during training, compelling the model to distinguish true MFs from a wide array of confounders and domain-specific artifacts.
The detection model is fundamentally based on RTMDet, a one-stage anchor-based detector specifically optimized for real-time performance. It was adapted for small object detection by employing an appropriate anchor size (64 × 64 pixels, corresponding to typical MF size at 40× magnification) and leveraging transfer learning from a COCO-pretrained checkpoint. Training was conducted on image patches of 1920 × 1280 pixels, ensuring ample context around candidate mitoses.
A critical element of their training methodology is a three-level, balanced sampling strategy. This approach ensures that each mini-batch used for training comprises a roughly equal number of patches from each of the six source datasets, preventing any single dataset or domain from disproportionately influencing the gradient updates. Additionally, it strives to draw patches evenly from different whole-slide images within each dataset, exposing the model to a diverse range of patients and tissue regions. Finally, at the patch level, the strategy maintains a balanced ratio, approximately 1:1, between MF and non-MF instances. This multi-tiered balancing effectively mitigates inherent class imbalance and prevents “easy” negative examples from overwhelming the training process, thereby enhancing consistent performance across all domains.
Data augmentations were also judiciously applied to expose the model to minor variations in appearance and orientation, thereby improving its ability to handle differences in staining and imaging. Each training patch underwent random horizontal and vertical flips (50% probability) and a small random rotation (uniformly sampled from -15° to +15°). These geometric augmentations account for the fact that MFs can appear in any orientation. Furthermore, color jittering was performed in the HSV color space, randomly adjusting hue, saturation, and value within limited ranges. This simulates variations in staining (e.g., H&E dye concentration, scanner white balance) without altering the underlying tissue morphology. The researchers avoided excessive or unrealistic augmentations, as preliminary experiments indicated that preserving histological details was crucial for accurate MF detection.
The targeted hard negative mining involved applying their trained RTMDet-S model to public datasets containing necrotic or debris-labeled regions. During this process, a deliberately low detection confidence threshold of 0.25 was set to capture a broad set of candidate detections, including many spurious objects. All predictions located in necrotic regions or tissue debris, which by definition cannot contain true MFs, were labeled as non-MF examples and added to the training pool as hard negatives. This strategy exposed the model to a wide range of mitosis-like imposters (e.g., fragmented nuclei, condensed chromatin remnants, or irregular debris structures) and taught it to suppress confident predictions in such regions. Incorporating these mined negatives proved particularly effective in reducing false positives on challenging test slides containing necrosis, thereby enhancing overall generalization across domains.
The evaluation demonstrated consistently high detection quality, with F1 scores above 0.78 across all folds in the cross-validation. On the MIDOG 2025 preliminary leaderboard, their approach achieved an F1 score of 0.81, with a recall of 0.84 and precision of 0.78, positioning it among the top entries. A significant advantage of their method is its real-time inference capability: on an NVIDIA 3090 GPU, the RTMDet-S model processes a 1920 × 1280 image patch in approximately 20–30 milliseconds, equating to over 30 patches per second. Interestingly, experiments with larger versions of the detector (RTMDet-M, L, X, or variants using a Swin-L Transformer or ConvNeXt backbone) did not yield meaningful improvements and sometimes even resulted in slightly worse F1 scores. This suggests that increased model complexity led to overfitting the training domains despite regularization, adding computational cost without clear benefits for this specific task. The small model, therefore, struck the best balance between bias and variance.
Also Read:
- Unlocking Deeper Insights in Cancer Pathology with EAGLE-Net’s Spatial Learning
- Brain Tumor Detection Enhanced by a New Dual-Stream AI Model
In conclusion, this study effectively demonstrates that a compact one-stage detector, when combined with appropriate training techniques, can achieve excellent generalization in mitotic figure detection across a wide array of domains. This real-time capability is crucial for clinical deployment, where gigapixel-sized whole-slide images necessitate rapid processing of numerous patches. The findings pave the way for integrating automated MF detection into the pathologist’s workflow, offering quick second opinions or triaging slides based on proliferative activity, ultimately improving the objectivity and reproducibility of mitosis quantification and tumor grading. For more details, you can refer to the full research paper available at arXiv.
