Baseline ¹⁸F-FDG PET/CT habitat radiomics versus dual-channel deep learning for predicting interim PET early...

To develop baseline ¹⁸F-FDG PET/CT habitat radiomics and dual-channel deep learning models for the prediction of early metabolic response (EMR) on interim PET (iPET) in patients with DLBCL and compare their performance.

Patients with DLBCL who underwent baseline ¹⁸F-FDG PET/CT between December 2018 and August 2024 and received 2-4 cycles of R-CHOP or R-CHOP-like chemotherapy were retrospectively enrolled. Based on iPET Deauville scores, patients with scores of 1-3 were classified as EMR, whereas those with scores of 4-5 were classified as non-EMR. Lesions were semi-automatically segmented to generate volumes of interest (VOIs). Voxel-wise habitat subregions were delineated using K-means clustering to characterize intratumoral heterogeneity. Radiomics features were extracted from whole-tumor and habitat subregions on PET and CT images. For deep learning, the largest axial tumor slice from PET and CT was concatenated along the channel dimension as dual-channel input. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), accuracy, balanced accuracy, sensitivity, specificity, and F1 score. Calibration curves, decision curve analysis, net reclassification improvement, and integrated discrimination improvement were additionally used for model comparison.

A total of 148 patients (101 EMR, 47 non-EMR) were randomly split into training (n = 103) and test (n = 45) sets. This retrospective single-center study used the training set for feature selection and model development, and the test set exclusively for final evaluation. We developed habitat radiomics and dual-channel deep learning models; multilayer perceptron and DenseNet-161 were selected as the optimal architectures, respectively. In the test set, the habitat radiomics model (Habitat_MLP) achieved an AUC of 0.871 (95% CI: 0.7563-0.9857), with a specificity of 0.903, and accuracy of 0.822, showing higher overall performance than the dual-channel deep learning model (DL_DenseNet161), which achieved an AUC of 0.793 (95% CI: 0.6409-0.9444), specificity of 0.677, and accuracy of 0.711. Furthermore, the Habitat_MLP model showed more favorable calibration, higher net benefit on DCA, and improved risk reclassification in this dataset.

The habitat radiomics model derived from baseline ¹⁸F-FDG PET/CT (Habitat_MLP) demonstrated superior overall performance and robustness for predicting EMR on iPET in patients with DLBCL, suggesting its potential value as a decision-support tool for pretreatment risk stratification.