PET/CT Radiomics Combined with Machine Learning Predicts Outcomes and Key Mutations in CLL

A new study published in Diagnostics highlights how radiomic analysis of [18F]-FDG PET/CT scans, paired with machine learning (ML), can help predict disease progression and genetic mutations in patients with chronic lymphocytic leukemia (CLL). The findings support the potential of advanced imaging biomarkers as a noninvasive tool for risk stratification.

In this retrospective analysis of 50 patients newly diagnosed with CLL (32 men and 18 women, aged 60–84), researchers extracted 865 radiomic features from PET/CT scans taken at the time of diagnosis at Policlinico Tor Vergata in Rome, Italy. Lymph nodes showing suspected involvement were segmented for detailed analysis. All participants also underwent genetic testing for mutations in TP53, NOTCH1, and IGVH.

The patient group represented a broad range of disease severity based on RAI and Binet staging. For model development, patients were randomly split into training (70%) and validation (30%) sets. Separate feature sets were created for each imaging modality—PET and CT—tailored to predict disease progression or specific mutations.

Out of the entire cohort, 10 patients experienced disease progression. Random Forest models using wavelet-based radiomic features delivered the strongest predictive performance. For progression, the model achieved an area under the curve (AUC) of 0.94 in training and 0.88 in validation, with accuracy scores of 0.87 and 0.75, respectively.

When predicting TP53 mutations, PET-derived features led to an AUC of 0.96 and an accuracy of 0.80, with true positive and true negative rates over 90%. NOTCH1 mutation prediction also performed well, with a PET AUC of 0.85 and accuracy of 0.67. True positive and negative rates were 88.9% and 83.3%, respectively.

However, predictions for IGVH mutational status were less reliable. The Stochastic Gradient Descent model performed slightly better than Random Forest, but overall accuracy remained moderate (AUC >0.75). Notably, traditional clinical and volumetric PET/CT measures did not significantly differ among subgroups (P >.05), further emphasizing the unique value of radiomic features.

The researchers concluded that integrating radiomics with ML could offer new insights into CLL biology and prognosis.

“Our findings provide initial evidence correlating PET and/or CT functional parameters with the mutational profile of CLL patients,” they noted. “This first radiomic study of chronic lymphoproliferative syndromes suggests the potential of radiomics in identifying patients with poorer prognosis. Future integration of biological and radiological parameters could improve CLL risk stratification and inform clinical trial design.”

This study points toward a future where imaging not only helps diagnose disease but also guides personalized treatment decisions in hematologic cancers like CLL.