Exploring Multi-Pathology Brain Segmentation: From Volume-Based to Component-Based Deep Learning Analysis.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• AI algorithms are increasingly used for detecting and segmenting brain abnormalities in MRI scans.
• While AI models achieve high accuracy, in-depth evaluation of their failures is crucial for clinical integration.
• Key factors for evaluation include abnormality position, intensity, and volume.

Purpose of the Study:

• To analyze the segmentation performance of a pre-trained U-net model on brain MRI scans with four pathologies.
• To evaluate the model's accuracy in segmenting both whole abnormal volumes and individual abnormal components.
• To investigate the relationship between segmentation errors (True Positives, False Negatives, False Positives) and abnormality characteristics.

Main Methods:

• Utilized a pre-trained U-net model on a validation set of brain MRI scans.
• Assessed segmentation performance using Dice Score Coefficient (DSC), sensitivity, and precision for whole abnormal volumes.
• Analyzed the detection and segmentation of individual abnormal components, categorizing results as correct, partial, missed, or false positives.

Main Results:

• For whole abnormal volumes, the model achieved a DSC of 0.76, sensitivity of 0.78, and precision of 0.82.
• For individual abnormal components, 48.8% were correctly segmented (DSC ≥ 0.5), 27.1% partially segmented (0.05 > DSC > 0.5), and 24.1% missed (False Negatives).
• The model produced 25.1% False Positives, with further analysis correlating errors with abnormality position, intensity (FLAIR, T2, T1ce), and volume.

Conclusions:

• The U-net model demonstrates significant capability in segmenting brain abnormalities but requires further refinement for precise clinical application.
• Understanding failure modes related to abnormality characteristics is essential for improving AI model reliability in neuroimaging.
• Detailed analysis of segmentation errors provides insights for developing more robust AI tools for radiological practice.