A 3D Convolutional Neural Network Based on Non-enhanced Brain CT to Identify Patients with Brain Metastases.

Area of Science:

• Radiology
• Artificial Intelligence
• Oncology

Background:

• Non-enhanced CT (NE-CT) brain scans are increasingly common in cancer staging but often struggle to detect brain metastases (BM).
• Dedicated brain imaging is rarely recommended for asymptomatic cancer patients, creating a diagnostic gap.
• Developing AI tools for NE-CT could improve BM detection efficiency.

Purpose of the Study:

• To develop and validate a 3D Convolutional Neural Network (3D-CNN) for detecting brain metastases (BM) using non-enhanced CT (NE-CT) brain scans.
• To assess the accuracy of the 3D-CNN in differentiating cancer patients with and without BM.
• To establish a potential AI-driven method for early BM diagnosis without contrast enhancement.

Main Methods:

• Retrospective analysis of NE-CT scans from 100 patients with BM and 100 without, excluding lesions smaller than 5 mm.
• Manual segmentation of the largest tumor on contrast-enhanced MRI and extraction of radiomic features.
• Development and optimization of a 3D-CNN model, including convolutional layers, pooling, dropout, and sigmoid activation, trained on 70% and validated on 30% of the dataset.

Main Results:

• The best 3D-CNN model achieved a mean validation accuracy of 0.983 (SD: 0.020) and an area under the ROC curve of 0.983 (SD: 0.023).
• The model demonstrated high sensitivity (0.983, SD: 0.020) in detecting brain metastases.
• The study successfully developed an accurate AI model for BM detection on NE-CT scans, with a median largest tumor diameter of 2.29 cm.

Conclusions:

• A 3D-CNN model effectively differentiates between cancer patients with and without brain metastases using only non-enhanced CT scans.
• The developed AI tool shows high diagnostic performance, suggesting its potential utility in clinical settings.
• Further external validation is recommended to confirm the model's generalizability and clinical impact.