Contextual Deep Regression Network for Volume Estimation in Orbital CT.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Ophthalmology and neuroscience

Background:

• Optic nerve diseases cause structural changes detectable by computed tomography (CT).
• Multi-atlas methods segment optic nerve volumes, correlating with visual disability.
• Current methods for optic nerve volume estimation have limitations.

Purpose of the Study:

• To train a weakly supervised convolutional neural network (CNN) for direct optic nerve volume learning without segmentation.
• To investigate the impact of contextual electronic medical record (EMR) data, specifically ICD-9 codes, on optic nerve volume estimation.
• To compare the diagnostic performance of optic nerve volume biomarkers derived from different methods.

Main Methods:

• Trained a weakly supervised CNN to learn optic nerve volumes directly from CT images.
• Developed a merged network combining imaging data with EMR information (ICD-9 codes).
• Compared disease prediction models using volumes from multi-atlas segmentation, standard CNN, and contextual CNN.

Main Results:

• The merged network incorporating EMR context improved optic nerve volume prediction, showing a 15% increase in explained variance.
• The contextual CNN (merge-CNN) achieved an AUC of 0.74 for disease classification.
• This significantly outperformed the multi-atlas method (AUC of 0.54).

Conclusions:

• Contextual data from EMR significantly enhances optic nerve volume estimation accuracy.
• A contextually derived volume biomarker from CNNs is more accurate for disease prediction than traditional methods.
• Integrating non-imaging data holds substantial potential for improving medical image analysis and diagnostic accuracy.