Non-Hodgkin's lymphoma classification using 3D radiomics machine learning models for precision imaging in oncology

Area of Science:

• Radiology and Medical Imaging
• Computational Pathology
• Oncology

Background:

• Accurate classification of Non-Hodgkin Lymphoma (NHL) subtypes is crucial for predicting clinical behavior and guiding treatment strategies.
• Imaging plays a role in disease staging, and quantitative analysis of imaging features shows potential for predicting pathology and patient outcomes.
• Machine learning applied to imaging features offers a promising avenue for enhancing clinical decision-making in lymphoma diagnosis.

Purpose of the Study:

• To apply quantitative imaging analysis for noninvasive classification of frequent Non-Hodgkin Lymphoma (NHL) subtypes.
• To establish a foundation for a clinical imaging-genomic model to support therapeutic monitoring and clinical decision-making.
• To differentiate between lymphoma subtypes and non-lymphoma tissues, as well as among major NHL subtypes.

Main Methods:

• A retrospective analysis of 201 treatment-naïve NHL patients and 39 healthy controls using contrast-enhanced CT scans.
• Three-dimensional segmentation and radiomic analysis of 1,628 pathologically enlarged lymph nodes, with healthy lymph nodes serving as references.
• Feature selection using a random forest (RF) classifier, followed by multiclass classification using a Light Gradient Boosting Machine (LGBM) for lymphoma subtype differentiation.

Main Results:

• High accuracy was achieved in classifying lymphoma versus non-lymphoma (AUC=0.999).
• Excellent performance was observed in differentiating lymphoma subtypes: MCL vs. others (AUC=0.997), DLBCL vs. others (AUC=0.971), CLL vs. others (AUC=0.956), and FL vs. others (AUC=0.892).

Conclusions:

• Radiomics combined with multiclass machine learning enables highly accurate, non-invasive differentiation of major NHL subtypes on routine contrast-enhanced CT.
• This approach can reliably separate indolent from aggressive lymphoma phenotypes, paving the way for imaging-genomic models.
• The findings support streamlined biopsy guidance, enhanced therapeutic monitoring, and advancements in precision oncology for lymphoma care.