Anatomical context improves deep learning on the brain age estimation task.

Area of Science:

• Medical Imaging Analysis
• Artificial Intelligence in Healthcare

Background:

• Deep learning models excel at medical image analysis without manual feature engineering.
• Traditional methods rely on handcrafted structural imaging features.
• Integrating both approaches may enhance predictive accuracy.

Purpose of the Study:

• To investigate if deep learning is complementary to traditional feature estimation in medical imaging.
• To propose a network design combining deep convolutional and structural imaging features.
• To evaluate this approach for imaging-based age prediction.

Main Methods:

• A novel network architecture was designed to integrate raw images with engineered structural features.
• The approach was tested on T1-weighted brain MRI (N=5121) and head CT (N=1313) datasets.
• Age prediction accuracy was assessed using mean absolute error (MAE) and median absolute error (MedAE).

Main Results:

• In brain MRI, the combined approach achieved an MAE of 4.08 years, outperforming image-derived (5.00 years) and structural features (8.23 years) alone.
• For head CT, the combined method yielded a MedAE of 9.99 years, superior to image-derived (11.02 years) and structural features (13.28 years) alone.
• The results demonstrate improved age prediction by synergizing deep learning and traditional features.

Conclusions:

• Deep learning can effectively complement traditional feature estimation in medical image analysis.
• Combining deep learning with structural features enhances the performance of age prediction tasks.
• Future medical image processing should leverage deep learning to augment, not replace, traditional imaging features.