Featured article
Can AI Sharpen DLBCL Risk Prediction?
Risk assessment plays an important role in the management of diffuse large B-cell lymphoma, yet traditional clinical scoring systems do not capture the full biological complexity of the disease. PET/CT imaging provides important information about disease burden, while advances in genomic profiling are revealing molecular differences that may also influence prognosis. Researchers are now investigating whether artificial intelligence can help bring these different sources of information together to identify patients at greater risk of poor outcomes more precisely.
In a study of 1,024 newly diagnosed patients with DLBCL, investigators used a deep learning framework to analyze baseline FDG-PET/CT scans and automatically identify and quantify lymphoma lesions. The approach extracted several imaging biomarkers, including total metabolic tumor volume and measures of tumor dissemination. The automated measurements showed strong agreement with manually derived measurements, suggesting that deep learning may help address some of the complexity and variability associated with extracting quantitative biomarkers from whole-body PET imaging.
The researchers then combined PET-derived biomarkers with clinical characteristics and genetic information to develop the ClinicalPET LymphPlex prognostic index. High total metabolic tumor volume, elevated lactate dehydrogenase, and several molecular subtypes emerged as significant prognostic factors. Using these characteristics, the model separated patients into low-, low-intermediate-, high-intermediate-, and high-risk groups, with substantial differences in progression-free and overall survival observed across the groups.
The integrated model also demonstrated stronger prognostic discrimination than the International Prognostic Index in the study populations. Its prognostic value was subsequently evaluated in external and treatment-specific cohorts, where the model continued to distinguish patients with different outcomes. Additional analyses connected some of the model's risk factors with underlying tumor biology, including associations between high metabolic tumor volume and an immunosuppressive tumor microenvironment and between elevated LDH and increased tumor proliferation and metabolic activity.
These findings illustrate the potential of combining artificial intelligence, routinely acquired imaging, clinical information, and molecular profiling to refine DLBCL risk assessment. The approach remains investigational, and the researchers noted that larger independent cohorts and further improvements in classification accuracy are needed before broader clinical application. Still, integrating multiple dimensions of disease biology may eventually help clinicians move beyond conventional risk scores toward more precise prognostic assessment and, potentially, more individualized treatment strategies.