Methodological framework for radiomics applications in Hodgkin's lymphoma

Area of Science:

• Radiomics and Medical Imaging
• Oncology and Hematology
• Computational Pathology

Background:

• Radiomics features in Hodgkin's lymphoma (HL) vary between refractory/relapsing and long-term responders.
• Methodological aspects of radiomics in HL require further elucidation.

Purpose of the Study:

• Establish a radiomics methodological framework for Hodgkin's lymphoma (HL).
• Develop a novel feature selection approach for HL radiomics.
• Evaluate intra-patient lesion similarity and classify relapsing refractory (R/R) vs. non-(R/R) HL patients.

Main Methods:

• Retrospective analysis of 85 Hodgkin's lymphoma (HL) patients using LIFEx software for [18F]FDG-PET/CT segmentation and feature extraction.
• Feature selection based on correlation with volume; Principal Component Analysis (PCA) for feature transformation.
• Lesion similarity assessed using silhouette scores; classification of R/R vs. non-R/R patients using Random Undersampling Boosting of Tree Ensemble (RUBTE) with single (fingerprint_One) and all (fingerprint_All) lesion fingerprints.

Main Results:

• Intra-patient HL lesion similarity analysis showed low similarity (mean/median silhouette < 0.5), particularly in the non-R/R group.
• Classification accuracy for R/R vs. non-R/R patients was 62% using fingerprint_One (78% sensitivity, 53% specificity).
• Classification accuracy improved to 82% using fingerprint_All (70% sensitivity, 88% specificity).

Conclusions:

• Hodgkin's lymphoma (HL) lesions exhibit heterogeneity in radiomics signatures within patients, indicating random target lesion selection is inappropriate.
• The developed lesion similarity analysis highlights the importance of considering all lesions for accurate radiomics applications in HL.
• Utilizing all lesions in the classifier significantly improved the prediction of relapsing refractory (R/R) vs. non-(R/R) status.