Characterization of adrenal glands on computed tomography with a 3D V-Net-based model

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Adrenal lesions can exhibit inter-observer variability in routine diagnostic workflows.
• Accurate characterization of adrenal glands is crucial for patient management.

Purpose of the Study:

• To evaluate a 3D V-Net-based segmentation model for adrenal lesion detection.
• To assess the model's performance in classifying adrenal glands as normal or abnormal.
• To compare the model's performance against radiology reports.

Main Methods:

• Retrospective collection and annotation of 1086 CT image series with adrenal lesions for model training.
• Evaluation of segmentation performance using Dice Similarity Coefficient (DSC) on a test set.
• External validation on two cohorts (959 patients with confirmed lesions, 479 patients with malignancy history) assessing classification performance (sensitivity, accuracy).

Main Results:

• The segmentation model achieved a DSC of 0.900 on the test set.
• Sensitivities and accuracies were 99.7% and 98.3% for external validation dataset 1, and 87.2% and 62.2% for external validation dataset 2.
• The model's performance showed no significant difference compared to radiology reports in validation datasets.

Conclusions:

• The developed 3D V-Net model effectively segments adrenal lesions.
• The model can be utilized for binary classification of adrenal glands, aiding in abnormality detection.
• The model demonstrates high accuracy in pre-surgical scenarios and high sensitivity in screening populations.