CT-IDP: Segmentation-Derived Quantitative Phenotypes for Interpretable Abdominal CT Disease Classification

Technical Summary: CT-IDP – Segmentation-Derived Quantitative Phenotypes for Interpretable Abdominal CT Disease Classification

Problem Statement
The increasing volume of abdominal CT scans necessitates robust, automated computer-aided diagnosis methods that generalize across diverse scanners, protocols, and institutions. While deep learning models (CNNs, vision transformers, and vision-language models) have achieved strong benchmark performance, they often rely on "shortcut learning," prioritizing texture-dominant appearance features over stable, quantitative manifestations of disease. This limitation is particularly critical for radiologic findings defined by measurable anatomy—such as organ size, vessel caliber, tissue attenuation, or calcific burden—where disease expression is quantitative rather than purely textural. Furthermore, traditional radiomics approaches are typically lesion-centric and sensitive to reconstruction kernels and noise, lacking a unified framework for cross-disease application.

Methodology
The authors propose CT-IDP (CT Image-Derived Phenotypes), a framework that operationalizes an anatomy-centric formulation for multi-label abdominal CT classification. The methodology proceeds through four key stages:

1. Segmentation and Descriptor Extraction:

   • Multi-organ segmentations are generated using TotalSegmentator and mapped to the native CT image grid.
   • Over 900 organ and compartment-level descriptors are computed in native image space using voxel-spacing metadata to ensure physical units.
   • These descriptors are categorized into three families:
     • Morphometry: Volume, diameter, surface area, and shape ratios.
     • Attenuation: First-order HU summaries, cross-organ contrasts, and burden estimates.
     • Contextual/Burden: Cavity-normalized occupancy and composite ratios.

2. Sparse Disease-Specific Modeling:

   • Instead of a single end-to-end deep network, a separate sparse logistic regression model is trained for each target finding using an elastic-net regularizer.
   • Features are selected via stratified 5-fold cross-validation on the MERLIN development set. Highly correlated descriptors are filtered, and missing values are imputed using dataset summary statistics.
   • The final model specification (descriptors, imputation rules, scaling parameters, and coefficients) is "locked" and applied identically to external datasets without retraining.

3. Benchmarking:

   • CT-IDP is compared against a DINOv3-based vision transformer (ViT) baseline.
   • The ViT baseline is trained on the same MERLIN development data, using intensity clipping, resampling, and a tri-slice encoding scheme to aggregate slice-level embeddings into study-level predictions.

4. Evaluation Strategy:

   • Datasets: Development on the public MERLIN benchmark (15,175 training, 5,018 validation, 5,082 test studies). External validation on a private institutional dataset (n=2,000) and the public AMOS dataset (n=1,107).
   • Labels: Study-level labels are derived from radiology reports (using LLM-assisted NLP for the private dataset) and mapped to the phenotype library.
   • Metrics: Macro-AUC and Average Precision (AP) with 95% confidence intervals via bootstrapping.

Key Results

• Overall Performance: CT-IDP outperformed the DINOv3 baseline across all three datasets on macro-AUC and macro-AP.
  • MERLIN: CT-IDP AUC 0.897 vs. ViT 0.880.
  • Private Institutional: CT-IDP AUC 0.877 vs. ViT 0.857.
  • AMOS: CT-IDP AUC 0.780 vs. ViT 0.756.
• Regime-Specific Performance:
  • CT-IDP Superiority: The quantitative model demonstrated significant gains on findings defined by measurable anatomy or attenuation, including gallstones (+0.111 to +0.255 ΔAUC), hepatic steatosis (+0.157), abdominal aortic aneurysm (AAA), and splenomegaly.
  • ViT Superiority: The vision transformer outperformed CT-IDP on findings dependent on spatial patterns, compartmental fluid distribution, or bowel configuration, such as free air, bowel obstruction, and ascites.
  • Observability Limits: Performance analysis revealed that CT-IDP's discrimination drops sharply for sub-threshold or non-calcified lesions (e.g., non-calcified gallstones, small renal cysts) where the abnormality falls below the resolution of the segmentation-derived descriptors. In these regimes, the ViT retained moderate residual discrimination based on local appearance cues.
• Interpretability: The sparse models provided explicit, auditable prediction rules. Coefficient inspection confirmed that decision rules aligned with clinical thresholds (e.g., probability rising with aortic diameter and falling with liver-spleen attenuation difference). Failures were mechanistically inspectable, often traceable to specific descriptor limitations or label-provenance issues (e.g., NLP-driven co-labeling of hepatomegaly and steatosis).

Significance and Claims
The paper posits that CT-IDP defines a distinct and complementary operating regime to learned image representations:

• Measurement-Grounded Robustness: For diseases expressed through measurable anatomy (size, attenuation, burden), models built directly on these measurements are more robust to dataset shifts and more interpretable than appearance-based embeddings.
• Interpretability by Design: Unlike post-hoc explainability methods, CT-IDP's interpretability is intrinsic; every prediction is a sparse function of named anatomical measurements, allowing clinicians to inspect, contest, or audit specific decision rules.
• Failure Analysis: The framework facilitates the dissection of failure mechanisms. For instance, the paper identifies that external performance gaps in hepatomegaly were likely driven by NLP-derived training labels that conflated hepatomegaly with steatosis, a nuance revealed through coefficient inspection.
• Future Direction: The authors suggest a hybrid approach as the natural next step, combining explicit CT-IDP measurements with anatomy-aware visual embeddings, utilizing each in its respective regime (measurement-defined vs. spatial-pattern-defined).

Limitations Acknowledged
The study notes that descriptor fidelity depends on segmentation quality, particularly for small lesions. The use of report-derived labels compresses radiologically heterogeneous diseases into binary states, potentially obscuring operating regimes. Additionally, the lack of per-study contrast-phase metadata in external datasets limited the ability to isolate protocol variation as a specific cause of performance degradation. The sparse logistic regression approach, while transparent, may under-represent complex nonlinear interactions.