Enterprise AI Analysis
Normal and Atypical Mitosis Image Classifier using Efficient Vision Transformer
A deep dive into the MIDOG 2025 challenge solution, leveraging a hybrid CNN-ViT architecture for highly accurate and efficient mitosis classification across diverse cancer types.
Authored by: Xuan Qi, Dominic Labella, Thomas Sanford, Maxwell Lee | Publication: arXiv:2509.02589v1 [eess.IV] 28 Aug 2025
Executive Impact: Precision in Cancer Diagnostics
The EfficientViT model delivers high-performance, critical for integrating advanced AI into clinical pathology workflows for improved diagnostic accuracy and efficiency.
0
Balanced Accuracy (Preliminary Eval)
0
ROC AUC (Preliminary Eval)
0
Unified Dataset Size
0
Cancer Types Covered
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
EfficientViT: The Hybrid Powerhouse for Clinical AI
The model employs EfficientViT-L2, a hybrid CNN-ViT architecture designed for both accuracy and efficiency, critical for clinical deployment. With a 256x256 input, it achieves 85.37% ImageNet accuracy while maintaining low computational cost (64M parameters, 9.1 GMACs). Its cascaded linear attention mechanism captures long-range dependencies efficiently, making it ideal for complex image classification tasks like mitosis detection.
Robust Generalization: Leave-One-Out Cross-Validation & Ensemble
Exclude one cancer type
→
Train on 6 types (5-fold CV)
→
5 models per configuration
→
Evaluate on held-out type
→
Ensemble predictions
→
Final Model Submission
| Cancer Type | Sample Num | AMF | NMF | AMF % |
|---|---|---|---|---|
| Canine cutaneous mast cell tumor | 2327 | 351 | 1976 | 15.10% |
| Canine lung cancer | 855 | 110 | 745 | 12.90% |
| Canine lymphoma | 3959 | 317 | 3642 | 8.00% |
| Canine soft tissue sarcoma | 1286 | 210 | 1076 | 16.30% |
| Human breast cancer | 3722 | 832 | 2890 | 22.40% |
| Human melanoma | 1150 | 271 | 879 | 23.60% |
| Human neuroendocrine tumor | 639 | 85 | 554 | 13.30% |
| Total | 13938 | 2176 | 11762 | 15.60% |
Note: Class imbalance (average 15.6% AMF) was addressed with strategies like stain-deconvolution for augmentation and weighted loss functions.
0.942
Overall ROC AUC in Preliminary Evaluation
This metric highlights the strong discriminative capability of the model across all cancer domains in the preliminary evaluation phase.
| Domain | ROC AUC | Accuracy | Sensitivity | Specificity | BA |
|---|---|---|---|---|---|
| Domain 0 | 0.875 | 0.806 | 1.000 | 0.781 | 0.891 |
| Domain 1 | 0.898 | 0.820 | 0.724 | 0.841 | 0.783 |
| Domain 2 | 0.974 | 0.888 | 0.972 | 0.854 | 0.913 |
| Domain 3 | 0.972 | 0.895 | 1.000 | 0.889 | 0.944 |
| Overall | 0.942 | 0.850 | 0.873 | 0.844 | 0.859 |
| Cancer Type | ROC_AUC | BA | NMF_acc | AMF_acc |
|---|---|---|---|---|
| canine_cutaneous_mast_cell_tumor | 0.965 | 0.909 | 0.894 | 0.923 |
| canine_lung_cancer | 0.958 | 0.873 | 0.928 | 0.818 |
| canine_lymphoma | 0.905 | 0.802 | 0.926 | 0.678 |
| canine_soft_tissue_sarcoma | 0.951 | 0.879 | 0.892 | 0.867 |
| human_breast_cancer | 0.913 | 0.830 | 0.777 | 0.882 |
| human_melanoma | 0.951 | 0.876 | 0.863 | 0.889 |
| human_neuroendocrine_tumor | 0.969 | 0.894 | 0.894 | 0.894 |
Calculate Your Potential AI Impact
Estimate the direct ROI of implementing advanced AI solutions in your enterprise, based on efficiency gains and cost reductions.
Estimated Annual Savings
$0
Hours Reclaimed Annually
0
Your AI Implementation Roadmap
A typical phased approach to integrate advanced AI solutions into your existing enterprise infrastructure and workflows.
Phase 1: Discovery & Strategy (2-4 Weeks)
Detailed assessment of current mitosis detection workflows, identification of integration points for the EfficientViT model, and definition of custom requirements.
Phase 2: Data Preparation & Model Adaptation (6-12 Weeks)
Establish secure data pipelines, perform necessary data annotation and augmentation tailored to your specific histological data, and fine-tune the EfficientViT model for optimal performance in your environment.
Phase 3: Integration & Testing (4-8 Weeks)
Seamless integration of the AI model with existing pathology systems (e.g., LIS, PACS), comprehensive testing, and initial user acceptance trials to ensure accuracy and workflow compatibility.
Phase 4: Deployment & Monitoring (Ongoing)
Full-scale deployment of the AI-powered mitosis classifier, continuous performance monitoring, and iterative improvements based on real-world clinical feedback and new data.
Ready to Transform Your Diagnostic Capabilities?
Connect with our AI specialists to discuss how EfficientViT and similar advanced models can enhance precision and efficiency in your pathology lab.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat