Detection of adrenal gland lesions on CT: Development and external validation of a deep-learning-based model.

OBJECTIVES

We wanted to develop an AI model for the detection of adrenal lesions on CT scans and validate the model on internal and external datasets.

MATERIALS AND

METHODS

The model was trained on 647 CT-scans covering the upper abdomen with at least one adrenal lesion. In the internal test dataset 142 consecutive adult patients (median 72 years, 56% female) with adrenal lesions were retrospectively included, while the external dataset included 50 patients with known lung tumors and adrenal lesions. Presence of a lesion was the reference and the prediction score by the model was the index. We evaluated sensitivity and specific at prespecified thresholds and performed ROC-analysis. We also tested for the influence of lesion size, sex, age, scanner type and scan protocol on accuracy using Fisher's exact test and did a false positive / negative analysis.

RESULTS

The model had a sensitivity and specificity in the internal test cohort of 91% (86%-95%) and 92% (86%-97%) and 95% (85%-99%) and 94% (81%-99%) in the external cohort. AUROC ranged between 0.89 and 1.0. There was a significant (p<0.001) lower accuracy in lesions below 1cm and above 4cm. False positive findings were significantly more often on the left (p=0.027) and false negative findings significantly more often on the right (p=0.035). Sex, age, scanner type and scan protocol did not significantly affect accuracy.

CONCLUSION

We developed a model for the detection of adrenal lesions on CT scans covering the upper abdomen. The model achieved a high sensitivity and specificity in an internal and an external validation dataset.