Diffusion-Based Quality Control of Medical Image Segmentations across Organs

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Deep learning (DL) enables automated medical image segmentation for large-scale studies.
• DL segmentation methods can produce anatomically implausible results (hallucinations).
• Existing automated quality control (QC) methods are organ-specific and lack generalizability.

Purpose of the Study:

• To develop a novel, organ-agnostic automated quality control (QC) framework for deep learning-based medical image segmentation.
• To address the limitations of existing QC methods by creating a self-adapting and generalizable solution.
• To improve the reliability of automated analysis pipelines in medical imaging.

Main Methods:

• Proposed no-new Quality Control (nnQC), a framework utilizing a diffusion-generative paradigm.
• Introduced a novel Team of Experts (ToE) architecture combining 3D spatial awareness and anatomical visual features.
• Integrated fingerprint adaptation for cross-organ, cross-dataset, and cross-modality adaptability.

Main Results:

• nnQC demonstrated consistent outperformance over state-of-the-art methods across seven organs and fifteen datasets.
• The framework proved effective even with highly degraded or missing segmentation masks.
• nnQC showed versatility and effectiveness across diverse organs and imaging modalities.

Conclusions:

• nnQC provides a robust, self-adapting QC solution for deep learning-based medical image segmentation.
• The proposed framework overcomes the organ-specificity limitations of previous QC methods.
• nnQC enhances the reliability and scalability of automated medical image analysis pipelines.