Individualized and Generalized Learner Models for Predicting Missed Hepatic Metastases.

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Radiologist diagnostic performance in detecting liver metastases on CT scans varies significantly.
• This variation cannot be attributed to differences in CT acquisition protocols.
• Understanding reader-specific detection patterns is crucial for improving diagnostic accuracy.

Purpose of the Study:

• To investigate the use of deep convolutional neural networks (CNNs) to predict radiologist detectability of liver metastases.
• To compare the performance of generalized and individualized CNN models in predicting reader-specific lesion detection.
• To evaluate the utility of radiomic features in predicting lesion detectability.

Main Methods:

• A multi-reader-multi-case study involving 10 radiologists and 102 contrast-enhanced CT liver scans.
• Ground truth lesions were established using histopathology or tumor progression.
• Deep convolutional neural networks (CNNs) and random forests were trained to predict lesion detection by average or individual radiologists using image patches and radiomic features.

Main Results:

• Individualized CNN models achieved higher performance (AUC = 0.82) in predicting reader-specific detectability compared to generalized CNN models (AUC = 0.78).
• Predictors based on radiomic features showed limited ability to differentiate detected from missed lesions (individualized AUC = 0.64, generalized AUC = 0.59).
• CNNs demonstrated superior ability to learn features predicting reader detectability over radiomic features.

Conclusions:

• CNNs can effectively learn automated features that predict radiologist detectability of liver lesions.
• Individualized prediction models show promise for identifying difficult-to-detect lesions and enabling targeted radiologist training.
• Substantial training data per reader is necessary for effective individualized prediction models.