The significance of PET/CT combined with machine learning models for the classification of lymphoma involvement and metastases in enlarged lymph nodes

TL;DR

This study shows that combining PET/CT imaging with machine learning, especially Random Forest models, can accurately distinguish lymphoma from metastases in enlarged lymph nodes.

Contribution

The novel integration of PET/CT metabolic profiling with machine learning models improves classification of lymph node pathology.

Findings

Lymphomatous nodes showed higher SUVmax, larger size, and increased splenic metabolism compared to metastatic nodes.

The Random Forest model achieved 93.88% accuracy and 100% specificity in classifying lymph node involvement.

Splenic metabolic parameters significantly improved model performance in differentiating lymphoma from metastases.

Abstract

Accurate differentiation between lymphoma involvement and lymph node metastasis poses significant diagnostic challenges due to overlapping imaging characteristics. This study evaluates the discriminative capacity of PET/CT metabolic profiling integrated with machine learning for nodal pathology classification.

We analyzed 247 lymph nodes from patients with diffuse large B-cell lymphoma (DLBCL, n=39) and solid tumor metastases (n=46). Multivariable logistic regression identified key PET/CT biomarkers, including metabolic parameters and anatomical features. Three machine learning models—Random Forest (RF), Support Vector Machine (SVM), and Artificial Neural Network (ANN)—were trained using these predictors.

Lymphomatous nodes exhibited significantly elevated metabolic activity (SUVmax median: 16.0 vs. 10.0, P<0.001), larger short-axis diameters (13 mm vs. 11 mm, P<0.001), and concurrent splenic hypermetabolism (spleen SUVmax 3.1 vs. 2.8, P<0.001). The RF model demonstrated exceptional performance with an AUC of 0.942, accuracy of 93.88%, and 100% specificity, outperforming SVM (AUC = 0.850) and ANN (AUC = 0.824). Splenic metabolic parameters significantly enhanced model discrimination.

Integration of PET/CT-derived SUVmax and splenic metabolic features with machine learning, particularly RF algorithms, provides a potential framework for distinguishing lymphoma-involved from metastatic nodes. This approach holds promise for optimizing biopsy decisions and refining pretreatment risk stratification in clinical oncology.

Figures