TLDR: Researchers developed an EfficientViT-L2 based AI model to classify normal and atypical mitosis in cancer images for the MIDOG 2025 challenge. Using a unified dataset of 13,938 nuclei from seven cancer types and a leave-one-cancer-type-out cross-validation strategy, the model achieved strong performance with a balanced accuracy of 0.859 and ROC AUC of 0.942, demonstrating its potential for accurate and efficient pathology analysis.
Researchers have developed an advanced image classifier designed to distinguish between normal and atypical mitosis, a crucial step in cancer diagnosis and research. This work, presented by Xuan Qi, Dominic Labella, Thomas Sanford, and Maxwell Lee, addresses the challenges of accurately identifying atypical mitosis, which is often indicative of cancerous growth.
The team participated in the MIDOG 2025 challenge, focusing on creating a robust and efficient solution. They employed EfficientViT-L2, a sophisticated hybrid architecture that combines the strengths of Convolutional Neural Networks (CNNs) and Vision Transformers (ViTs). This combination is particularly effective because CNNs are good at capturing local features, while ViTs excel at understanding global context through their self-attention mechanisms. The EfficientViT-L2 model was chosen for its balance of high accuracy and computational efficiency, which is vital for processing the vast number of image patches found in whole-slide pathology images.
A significant aspect of this research involved compiling a comprehensive dataset. They unified two existing datasets, MIDOG and AMi-Br, into a single collection comprising 13,938 nuclei. These samples spanned seven different cancer types, including both canine and human cancers. A key challenge addressed was the inherent class imbalance within the dataset, where atypical mitoses (AMF) constituted only about 15% of the total samples. To counter this, strategies like stain-deconvolution for image augmentation were used, which helps the model learn from a broader range of staining conditions without altering tissue morphology.
Also Read:
- Enhancing Mitotic Figure Detection in Cancer Diagnostics with Attention-Guided AI
- Investigating Synthetic Image Augmentation for Atypical Cell Detection in Histopathology
Ensuring Robustness and Generalization
To ensure the model’s ability to perform well on unseen data and different cancer types, the researchers implemented a rigorous evaluation strategy. They used a “leave-one-cancer-type-out cross-validation” (LOOCV) approach. This meant training the model on six cancer types and testing its performance on the one type it had not seen during training. This method simulates real-world diagnostic scenarios where models might encounter new or varied cancer presentations. Additionally, for each LOOCV setting, a 5-fold cross-validation was applied, leading to an ensemble of models whose predictions were averaged for improved reliability.
The model’s performance was evaluated using several metrics. In the preliminary evaluation phase of the MIDOG 2025 challenge, the ensemble model achieved a balanced accuracy of 0.859, a ROC AUC of 0.942, and a raw accuracy of 0.85. These figures demonstrate the model’s competitive and well-balanced performance across different measures, indicating its strong capability to distinguish between normal and atypical mitotic figures. The LOOCV tests also showed strong discriminative power with consistently high AUC values across most individual cancer types.
This research highlights a promising step forward in automated pathology, offering a tool that can assist in the accurate and efficient classification of mitosis, potentially aiding in cancer diagnosis and prognosis. For more detailed information, the full research paper can be accessed here.
