CT-IDP: Segmentation-Derived Quantitative Phenotypes for Interpretable Abdominal CT Disease Classification
A new research paper introduces CT-IDP, a quantitative phenotyping framework designed for interpretable abdominal CT disease classification. Developed through a retrospective multi-institutional study, the framework utilizes the MERLIN abdominal CT benchmark, comprising over 25,000 studies across training, validation, and test sets. The system generates multi-organ segmentations using TotalSegmentator to derive more than 900 descriptors related to morphometry, attenuation, and contextual findings. These features feed into sparse disease-specific logistic regression models with elastic-net regularization. The model's performance was externally validated on two independent datasets, Duke-Abdomen and AMOS, and compared against a DINOv3-based vision-transformer baseline. Results indicated that CT-IDP achieved superior Macro-AUC scores across all datasets: 0.897 versus 0.880 on MERLIN, 0.877 versus 0.857 on Duke-Abdomen, and 0.780 versus 0.756 on AMOS. This study highlights the potential of segmentation-derived phenotypes to enhance diagnostic accuracy and interpretability in medical imaging, offering a robust alternative to black-box deep learning models in clinical settings.
Wire timeline
CT-IDP: Segmentation-Derived Quantitative Phenotypes for Interpretable Abdominal CT Disease Classification
A new research paper introduces CT-IDP, a quantitative phenotyping framework designed for interpretable abdominal CT disease classification. Developed through a retrospective multi-institutional study, the framework utilizes the MERLIN abdominal CT benchmark, comprising over 25,000 studies across training, validation, and test sets. The system generates multi-organ segmentations using TotalSegmentator to derive more than 900 descriptors related to morphometry, attenuation, and contextual findings. These features feed into sparse disease-specific logistic regression models with elastic-net regularization. The model's performance was externally validated on two independent datasets, Duke-Abdomen and AMOS, and compared against a DINOv3-based vision-transformer baseline. Results indicated that CT-IDP achieved superior Macro-AUC scores across all datasets: 0.897 versus 0.880 on MERLIN, 0.877 versus 0.857 on Duke-Abdomen, and 0.780 versus 0.756 on AMOS. This study highlights the potential of segmentation-derived phenotypes to enhance diagnostic accuracy and interpretability in medical imaging, offering a robust alternative to black-box deep learning models in clinical settings.
cs.AI updates on arXiv.org