Predictive value of 18F-FDG PET/CT PyRadiomics features using a tabular deep learning model for cancer-related cognitive impairment in patients with DLBCL without CNS involvement

Abstract

Objective To evaluate the predictive value of tabular deep learning models trained on handcrafted 18F-FDG PET/CT PyRadiomics features and clinical data for cancer-related cognitive impairment (CRCI) among patients with diffuse large B-cell lymphoma without central nervous system involvement (non-CNS DLBCL). Material and methods We retrospectively analyzed 112 non-CNS DLBCL patients who underwent brain 18F-fluorodeoxyglucose (18F-FDG)PET/CT. Based on Montreal Cognitive Assessment (MoCA) and Functional Assessment of Cancer Therapy-Cognitive Function (FACT-Cog) scores, patients were classified into the CRCI group (n = 73) and the non-CRCI group (n = 39) and randomly partitioned into training (n=89) and test (n=23) sets at an 8:2 ratio. Five-fold cross-validation was performed within the training set for model development, and bootstrap resampling was used to estimate 95% confidence intervals for AUC. Radiomic features were extracted with PyRadiomics from automatically segmented brain regions using MOOSE120 software. These handcrafted PET/CT radiomic features and clinical variables were organized as structured tabular datasets; raw PET/CT images were not used as direct model inputs. Feature selection was performed using correlation coefficients. The performance of seven tabular deep learning models was evaluated using receiver operating characteristic (ROC) curves and decision curve analysis (DCA), and different input modalities (clinical, PET, CT, and combined) were compared. Shapley additive explanations (SHAP) were applied to interpret feature contributions. Results The AutoInt and Neural Oblivious Decision Ensembles (NODE) models showed the best predictive performance, achieving training areas under the curve (AUCs) of 0.905 and 0.904, and test AUCs of 0.858 and 0.808, respectively. DCA suggested potential clinical utility. In the modality-specific analysis, the AutoInt model integrating CT, PET, and clinical features achieved a training AUC of 0.905 and a test AUC of 0.858, representing the most favorable test-set performance among the evaluated AutoInt input modalities. A CT texture feature from the left occipital lobe emerged as the highest-ranked model contributor in SHAP analysis, with a mean absolute attribution score of approximately 0.0078. Conclusions Tabular deep learning models based on handcrafted PET/CT PyRadiomics features showed preliminary potential for predicting CRCI in patients with non-CNS DLBCL. Among the evaluated models, AutoInt achieved the most favorable performance in the exploratory internal test set. Multimodal integration of PET, CT, and clinical tabular features may provide complementary information for CRCI prediction. SHAP analysis identified a left occipital CT texture feature as the top-ranked model contributor; however, this finding should be regarded as exploratory rather than a validated biomarker.