Accelerating the search for antimicrobial peptides with Google Antigravity
deepmind.google·2h ago
AI Research
arXiv CS.AI·1d ago·5 min read
Beyond Discrimination: Evaluating Probability Calibration in Radiomics-Image Fusion for Renal Cell Carcinoma Classification
TL;DR
Radiomics-image fusion models for renal cell carcinoma classification were evaluated, revealing that these models significantly enhance both discrimination and probability calibration. The 3D MedVAE gated-fusion model showed the best performance across multiple metrics.
✦ Why It Matters
Consider implementing radiomics-image fusion models to improve classification accuracy in renal cell carcinoma.
Key Takeaways
- Integrate radiomics with image models for better classification.
- Focus on both discrimination and probability calibration metrics.
- Utilize the 3D MedVAE gated-fusion model for optimal performance.
- Evaluate multiple scoring measures to assess model effectiveness.
Full Summary
Reliable probability estimates are crucial for interpreting machine-learning classifiers in clinical settings, yet many models focus solely on discrimination metrics. This study assessed the impact of integrating radiomics with four image-model families using data from 396 patients in the KiTS23 dataset.
Seventeen individual models and four probability-average ensembles were evaluated based on discrimination, probability calibration, and scoring measures. Results showed that all 12 comparisons between image-only and fusion models improved in area under the curve (AUC) and average precision (AP), with Brier score and negative log likelihood (NLL) improving in 11 of 12 cases.
The 3D MedVAE gated-fusion model achieved the highest performance overall. The findings indicate that radiomics-image fusion can enhance probabilistic prediction quality, with calibration behavior varying across different architectures and fusion strategies.