Clinical Report: Enhanced Detection of Lymphoma Lesions in Whole-Body PET/CT
Overview
This study demonstrates that incorporating anatomical prior information significantly improves lesion detection performance in CNN-based models for lymphoma imaging using PET/CT. The findings highlight the limitations of transformer-based architectures in leveraging anatomical insights for enhanced detection.
Background
Lymphoma is a prevalent hematological malignancy characterized by diverse biological behaviors and clinical outcomes. Accurate detection of multiple lesions in lymphoma is critical for effective treatment and improved patient prognosis. The integration of 18F-FDG PET/CT imaging provides valuable metabolic and anatomical information, yet challenges remain in detecting all lesions, particularly in multifocal disease.
Data Highlights
Model
mAP@0.1–0.5
nnDetection (without anatomical masks)
0.288
nnDetection (with anatomical masks)
0.335
Key Findings
Incorporating anatomical priors improved lesion detection performance in the nnDetection framework.
nnDetection with anatomical masks showed a significant increase in mAP@0.1–0.5 from 0.288 to 0.335.
The Swin Transformer did not demonstrate clear advantages when anatomical priors were added.
Two independent PET/CT datasets were utilized, covering different tracers and cancer subtypes.
Explicit anatomical context is crucial for enhancing multi-lesion detection in CNN-based models.
Clinical Implications
The findings suggest that integrating anatomical insights can enhance the accuracy of lesion detection in lymphoma imaging, potentially leading to better clinical assessments. However, the limited effectiveness of transformer-based models indicates a need for further research in this area.
Conclusion
Incorporating anatomical priors into CNN-based models significantly enhances lesion detection in lymphoma imaging, while transformer-based approaches require improved strategies for effective integration.
