Discrepancy-based diffusion models for lesion detection in brain MRI.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Diffusion probabilistic models (DPMs) excel in image generation but require costly annotations for medical imaging.
• Existing DPM lesion detection methods rely on image-level annotations or original multi-modalities, limiting their effectiveness.

Purpose of the Study:

• To develop a novel DPM framework for brain MRI lesion detection that does not directly rely on image-level annotations or original modalities.
• To improve lesion detection performance by leveraging discrepancies in image-level annotations.

Main Methods:

• Proposed the Discrepancy Distribution Medical Diffusion (DDMD) model for brain MRI lesion detection.
• Translated image-level annotation inconsistencies into distribution discrepancies among heterogeneous samples.
• Preserved information within homogeneous samples to maintain pixel-wise uncertainty and enable implicit segmentation ensembles.

Main Results:

• The DDMD model demonstrated superior performance in brain lesion detection compared to state-of-the-art methods.
• Experiments were conducted on the BRATS2020 benchmark dataset for brain tumor detection using multimodal MRI scans.

Conclusions:

• The DDMD model offers a novel and effective approach to lesion detection in brain MRI.
• By utilizing annotation discrepancies, the model enhances detection performance and addresses limitations of current DPM-based methods.