Localization and Classification of Adrenal Masses in Multiphase Computed Tomography: Retrospective Study

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Adrenal incidentalomas are increasingly common, often requiring surgery.
• Accurate classification of adrenal masses from CT scans is challenging for clinicians, necessitating advanced tools.
• Current diagnostic methods can lead to increased workload and unnecessary surgeries.

Purpose of the Study:

• To develop an automated, noninvasive approach for adrenal mass identification and classification.
• To enhance diagnostic efficiency and transform preoperative diagnosis of adrenal masses.
• To improve the accuracy and reduce the subjectivity in classifying adrenal masses.

Main Methods:

• Retrospective analysis of adrenalectomy patient data from two centers (internal and external datasets).
• Development of a deep learning model, Multi-Attention YOLO (MA-YOLO), for localization and classification of 6 common adrenal mass types.
• Evaluation using intersection over union and mean average precision metrics, with comparison of diagnostic performance before and after model assistance.

Main Results:

• The MA-YOLO model achieved high performance in localization (IoU 0.838-0.890) and classification (mAP 0.885-0.915) on the external test set.
• Model assistance significantly improved diagnostic performance for radiologists and clinicians across most adrenal mass types.
• Notable improvements were observed in classifying adrenal adenoma and adrenal cortical carcinoma.

Conclusions:

• The MA-YOLO model offers an efficient, accurate, and noninvasive method for preoperative localization and classification of adrenal masses.
• This deep learning approach shows significant potential to aid clinical decision-making in adrenal mass diagnosis.
• The study highlights the value of AI in improving diagnostic accuracy and patient care for adrenal incidentalomas.