TLDR: Researchers from Xi’an Jiaotong-Liverpool University developed an adaptive AI framework for the MIDOG2025 challenge to classify atypical mitotic figures (AMFs), which are crucial indicators of abnormal cell division. Their approach, based on the UNI2-h pathology foundation model, uses visual prompt tuning (VPT) for efficient adaptation and integrates test-time augmentation (TTA) with stain normalization. This strategy significantly improved generalization and robustness, achieving a balanced accuracy of 0.8837 and an ROC-AUC of 0.9513, placing them in the top 10 teams.
Atypical mitotic figures (AMFs) are critical indicators of abnormal cell division, playing a significant role in assessing tumor proliferation and grading. However, their identification in routine histopathological slides is notoriously challenging due to their subtle, varied appearances and the inconsistencies introduced by different scanning equipment. Manual identification is time-consuming and prone to human error, while traditional deep learning methods often struggle with class imbalance, high intra-class variability, and the fine distinctions required for accurate classification.
The MIDOG2025 Track 2 challenge specifically addresses these difficulties by providing a benchmark for robust atypical mitotic figure classification under diverse imaging conditions. In response to this challenge, researchers Biwen Meng, Xi Long, and Jingxin Liu from Xi’an Jiaotong-Liverpool University developed an innovative framework to enhance the reliability of AMF detection.
Their approach builds upon UNI2-h, a powerful pathology foundation model known for its strong histopathological representations. To adapt this model for the complexities of atypical mitosis classification and scanner variability, the team introduced three key strategies:
Visual Prompt Tuning (VPT) for Efficient Adaptation
Instead of retraining the entire UNI2-h model, which can be computationally intensive, the researchers adopted Visual Prompt Tuning (VPT). This method involves inserting learnable prompt tokens before each transformer encoder block and updating only these tokens along with the classification head, while the main backbone parameters remain frozen. This significantly reduces the number of trainable parameters, allowing the model to efficiently capture the distinctive morphological patterns of atypical mitoses without extensive retraining.
Domain-Adversarial Learning for Scanner-Specific Biases
To ensure the model performs consistently across different scanners, a domain classifier was integrated into the shared feature space. This classifier is trained with a Gradient Reversal Layer (GRL) using scanner labels. The adversarial loss generated by this process penalizes the backbone if it retains scanner-specific information, thereby encouraging the learning of features that are invariant to the scanning device. This strategy is crucial for improving generalization to unseen scanners.
Also Read:
- Advancing Astrocyte Detection with a Hybrid AI Model
- Advancing Ulcerative Colitis Diagnosis with Lesion-Aware AI Captioning
Stain Normalization and Test-Time Augmentation (TTA) for Robustness
During inference, the framework employs a Test-Time Augmentation (TTA) strategy to boost prediction robustness. Each test image patch is evaluated under multiple transformations, including geometric augmentations like flips and rotations. Crucially, two complementary stain normalization pipelines, Vahadane and Macenko, are applied. These normalize the appearance discrepancies across samples, reducing variability caused by different staining protocols. The predictions from all augmented versions are then averaged to produce a final, more reliable probability.
The team systematically evaluated these variants on the MIDOG2025 Track 2 dataset. Their baseline, using LoRA (Low-Rank Adaptation) with UNI2-h, achieved a balanced accuracy of 0.8305 and an ROC-AUC of 0.9364. Replacing LoRA with Visual Prompt Tuning (VPT) significantly improved these metrics to a balanced accuracy of 0.8711 and an ROC-AUC of 0.9483. The most robust performance was achieved by further integrating Test-Time Augmentation (TTA) with Vahadane and Macenko stain normalization, which yielded a balanced accuracy of 0.8837 and the highest ROC-AUC of 0.9513.
These results demonstrate that a combination of prompt-based adaptation and stain-normalization TTA offers a highly effective strategy for robust atypical mitosis classification under diverse imaging conditions. The final submission ranked within the top 10 teams on the preliminary leaderboard, underscoring the practical utility of their adaptive learning strategies. For more details, you can read the full research paper here.
