Development and validation of a deep learning model using MR imaging for predicting brain metastases: an...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Brain metastases (BM) pose a significant threat to patients with extracranial malignancies.
• Manual identification of BM is labor-intensive and time-consuming.
• Accurate and efficient BM diagnosis is critical for patient outcomes.

Purpose of the Study:

• To develop and validate an automated system for diagnosing brain metastases.
• To assess the performance, stability, and clinical applicability of the developed system.

Main Methods:

• A 3D U-Net model with a ResNet-34 backbone was developed for BM prediction.
• MRI scans underwent preprocessing including resampling, skull stripping, and normalization.
• The model was trained and validated on a large dataset of 470 patients, including internal and external validation sets.

Main Results:

• The system demonstrated perfect specificity across all demographic groups.
• Performance varied by cancer type, with lower sensitivity for lung cancer compared to other cancers.
• The automated system significantly reduced diagnosis time by 40%-50% compared to manual annotation.

Conclusions:

• The developed AI system is robust, highly specific, and efficient for brain metastases diagnosis.
• Further optimization is needed to improve sensitivity for lung cancer cases.
• The system's efficiency and generalization capabilities support its potential for clinical translation.