Individualized and generalized models for predicting observer performance on liver metastasis detection using CT.

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Radiologist diagnostic performance shows significant inter-reader variability.
• Accurate detection of liver metastases in contrast-enhanced computed tomography (CT) is crucial for patient management.
• Developing predictive models can help understand and potentially mitigate diagnostic variability.

Purpose of the Study:

• To compare different feature sets (semantic, radiomic, CNN) for predicting radiologist liver metastasis detection.
• To evaluate the effectiveness of individualizing predictive models to specific radiologists.
• To assess potential improvements in diagnostic performance through personalized modeling.

Main Methods:

• Utilized abdominal CT images from 102 patients with 124 liver metastases.
• Generated 510 image sets by reconstructing images at different kernels/doses.
• Extracted features using logistic regression (semantic), random forests (radiomic), and convolutional neural networks (CNNs).
• Trained both generalized and individualized predictive models for metastasis detection.

Main Results:

• Individualized and generalized CNN models showed higher predictive accuracy (AUC) than semantic and radiomic models.
• Individualized CNN models (AUC 0.85 ± 0.04) outperformed generalized CNN models.
• Individualized semantic and radiomic models also surpassed their generalized counterparts.

Conclusions:

• Individualized models offered a slight performance advantage over generalized models across all feature sets.
• Inductive CNNs demonstrated superior performance in predicting metastasis detection compared to semantic or radiomic features.
• Generalized models offer practical advantages when individualized data are not accessible.