Unraveling Normal Anatomy via Fluid-Driven Anomaly Randomization.

Area of Science:

• Medical image analysis
• Machine learning
• Neuroimaging

Background:

• Current machine learning models for medical imaging are often modality-specific and struggle with variations in resolution and pathology.
• General-purpose models typically perform poorly on diseased subjects, limiting their clinical utility.

Purpose of the Study:

• To introduce UNA (Unraveling Normal Anatomy), a modality-agnostic approach for reconstructing normal brain anatomy.
• To develop a method that can handle both healthy and pathological brain scans without fine-tuning.
• To enable the use of general-purpose models on uncurated clinical images with pathology.

Main Methods:

• Developed a fluid-driven anomaly randomization technique to generate diverse pathology profiles.
• Trained UNA on a combination of synthetic and real medical imaging data.
• Validated the approach on healthy and stroke datasets using CT and MRI scans.

Main Results:

• UNA effectively reconstructs healthy brain anatomy across different modalities and resolutions.
• The model demonstrates direct applicability to anomaly detection in both simulated and real pathological scans.
• Showcased effectiveness on 3D healthy and stroke datasets, including CT and MRI.

Conclusions:

• UNA is the first modality-agnostic method for normal brain anatomy reconstruction that accommodates pathology.
• This approach bridges the gap between healthy and diseased image analysis, facilitating large-scale studies.
• UNA opens new avenues for analyzing uncurated clinical data in the presence of pathologies.