Clinical Knowledge-Guided PET/CT Lesion Segmentation With Interpretable Fusion of Metabolic and Structural Cues

Area of Science:

• Medical Imaging
• Radiology
• Artificial Intelligence in Medicine

Background:

• Automated whole-body lesion segmentation in 18F-FDG PET/CT is crucial for accurate tumor burden assessment in oncology.
• Manual segmentation suffers from interobserver variability, necessitating automated solutions.
• Current automated methods struggle with over/under-segmentation due to limited integration of metabolic and anatomical data.

Purpose of the Study:

• To develop a novel framework for refined PET/CT lesion segmentation by integrating anatomical and metabolic cues.
• To improve the accuracy and reproducibility of tumor segmentation in oncological diagnostics.
• To validate the prognostic significance of features extracted from the segmentation framework.

Main Methods:

• Development of a novel mixture-of-experts (MoE) based interpretable fusion module.
• Integration of complementary information from PET and CT modalities at the pixel level.
• Rigorous evaluation on three in-domain and two external datasets.

Main Results:

• Demonstrated superior segmentation performance and generalizability across multiple datasets.
• Visualizations provided insights into modality contributions to segmentation decisions.
• Validated the prognostic significance of extracted features for PET/CT-based prognosis predictions.

Conclusions:

• The proposed framework effectively refines PET/CT lesion segmentation by integrating multimodal information.
• The interpretable MoE module enhances understanding of modality contributions.
• The approach holds transformative potential for improving oncological diagnostics and prognosis prediction.