A CT-based deep learning model to differentiate between benign and malignant adrenal lesions.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Differentiating benign from malignant adrenal lesions is crucial for appropriate patient management.
• Current diagnostic methods can be limited, necessitating improved tools.

Purpose of the Study:

• To develop and evaluate a deep learning model for differentiating benign from malignant adrenal lesions using CT imaging.
• To assess the model's diagnostic performance by integrating radiological and non-radiological data.

Main Methods:

• A retrospective analysis of 385 pathologically confirmed adrenal lesions (101 malignant, 284 benign) from 380 patients.
• Development of four deep learning models using combinations of radiological (size, attenuation) and non-radiological (medical history, lab results) data.
• Model performance evaluated using sensitivity, specificity, accuracy, and AUC, with histopathology as the gold standard.

Main Results:

• Segmentation reproducibility achieved a Dice similarity coefficient of 0.92 ± 0.03.
• The most accurate model, integrating clinical, biological, and radiological data, achieved 84.2% accuracy and an AUC of 0.93 for diagnosing malignant adrenal lesions.
• High diagnostic capabilities were demonstrated on the test set.

Conclusions:

• A deep learning model integrating preoperative clinical, biological, and radiological features shows high capability in differentiating benign from malignant adrenal lesions.
• The developed model shows promise for improving the diagnostic accuracy of adrenal lesions on initial CT examinations.