TLDR: MitoDetect++ is a new deep learning pipeline designed for the MIDOG 2025 challenge, focusing on both mitosis detection and atypical mitosis classification in histopathological images. For detection, it uses an attention-augmented U-Net with an EfficientNetV2-L encoder. For classification, it employs a Virchow2 vision transformer fine-tuned with Low-Rank Adaptation (LoRA) for efficiency. The pipeline incorporates strong augmentations, focal loss, group-aware stratified cross-validation, and test-time augmentation (TTA) to enhance generalization and robustness against domain shifts. It achieves a balanced accuracy of 0.892, demonstrating its potential for clinical application.
In the crucial field of computational pathology, accurately identifying and classifying mitotic figures—especially distinguishing between normal and atypical ones—remains a significant challenge. These tasks are vital for precise tumor grading and predicting patient outcomes. The rarity of mitotic events, variations in how different observers interpret images, and a limited number of examples of atypical mitosis further complicate these efforts.
Addressing these complexities, researchers have introduced MitoDetect++, a comprehensive deep learning pipeline. This innovative system is specifically designed for the MIDOG 2025 challenge, tackling both the detection of mitosis (Track 1) and the classification of atypical mitosis (Track 2). The goal is to provide robust, adaptable solutions that can overcome domain variability and class imbalance often found in histopathological images.
Unpacking MitoDetect++: Mitosis Detection (Track 1)
For the first task, mitosis detection, MitoDetect++ employs an encoder-decoder model built with U-Net style skip connections. The core of its encoder is an EfficientNetV2-L architecture, which has been pre-trained on a vast image dataset (ImageNet). To enhance its ability to focus on relevant features, spatial and channel attention modules are integrated into the decoder blocks. This design allows the system to identify mitotic figures by treating their centroids as circular regions and performing a binary segmentation task.
The training process for detection involves a meticulous approach. The MIDOG++ dataset is initially split, with 10% reserved for testing and the remainder used for 5-fold cross-validation, ensuring stratification by tissue types to prevent data leakage. Patches of 512×512 pixels are extracted, and mitotic figure centroids are dilated to a diameter of 21 pixels, transforming detection into a segmentation problem. The model is trained using an AdamW optimizer and a combination of Jaccard, Dice, and Focal losses. A RandomSampler ensures that at least 40% of patches in each training batch contain mitotic figures, and early stopping prevents overfitting. For final predictions, an ensemble of the top three models from cross-validation is used, with their predictions averaged for increased robustness.
Unpacking MitoDetect++: Atypical Mitosis Classification (Track 2)
The second task, classifying atypical mitosis, leverages the powerful Virchow2 model, which is based on the Vision Transformer (ViT) architecture. This model is known for its multi-head self-attention mechanisms and its capacity to learn deep, hierarchical representations from images. For binary classification, its final classification head is modified to a single output neuron.
To make the fine-tuning process efficient and reduce computational demands, MitoDetect++ incorporates Low-Rank Adaptation (LoRA). LoRA is integrated into key components of the transformer, including the qkv projections, output projection layer, and both fully connected layers, significantly reducing the number of trainable parameters while maintaining model adaptability. A dropout rate of 0.3 is applied to the LoRA modules for better regularization.
Training for classification utilizes annotated mitotic patches from several datasets, including the MIDOG 2025 Atypical Set, AMi-Br Breast Cancer dataset, and others. Images are resized to 224×224 pixels. To prevent slide-level data leakage, a StratifiedGroupKFold approach is used, dividing data by slide identifiers while balancing atypical versus normal labels across folds. Extensive augmentations are applied during training, such as random resized crops, flips, rotations, color jitter, grayscale conversion, and random erasing, alongside StrongAugment, to enhance the model’s generalization capabilities. Validation transforms are kept simpler to mimic inference conditions.
The training configuration for classification includes an Adam optimizer, Focal Loss (which emphasizes tougher minority class examples), and a WeightedRandomSampler to counter label imbalance. Early stopping is also employed to halt training if validation improvement ceases. For inference, LoRA weights are merged into the base network, and Test-Time Augmentation (TTA) is applied, involving scaling, flips, rotations, and brightness adjustments. Predictions are then averaged across TTA variants and models for a robust final classification.
Also Read:
- Advancing Cancer Diagnosis: YOLOv12 for Automated Mitotic Figure Detection
- AI-Powered Classifier Advances Mitosis Analysis for Cancer Detection
Performance and Future Outlook
MitoDetect++ has demonstrated highly competitive overall performance, achieving a balanced accuracy of 0.892 across validation domains. While domains with many atypical examples showed near-perfect balanced accuracy, more challenging domains highlighted areas for further improvement, such as targeted strategies for sparse atypical instances. The research indicates that LoRA effectively reduces training overhead without sacrificing performance, and group-aware splits ensure that evaluations truly reflect generalization rather than memorization of specific slide patterns. Test-Time Augmentation and ensembling further boost robustness against domain shifts and input variability.
In conclusion, MitoDetect++ offers a resource-efficient and robust pipeline for detecting and subtyping atypical mitoses in histopathological images. Its integrated approach, combining LoRA-based fine-tuning, strong augmentation, and TTA-ensemble inference within a rigorously validated framework, makes it a strong candidate for future clinical and research deployment. You can read the full research paper here.
