BACKGROUND: To investigate the feasibility and accuracy of PET radiomics features, along with their combination with CT radiomics, dosiomics, and deep learning (DL) features, in predicting radiation pneumonitis (RP) in lung cancer patients treated with volumetric modulated arc therapy (VMAT). METHODS: A total of 206 and 27 lung cancer patients who underwent VMAT with pre-treatment PET/CT imaging were enrolled from Hospital One and Hospital Two for model training and external validation, respectively. Four machine learning (ML) methods were applied to build radiomics models with features extracted from CT (R_CT), PET (R_PET), radiomics features fused PET/CT (R_fFU) and fused PET/CT images (R_ iFU), as well dosiomics features (D). Three DL models were built to extract features from PET (DL_PET), CT (DL_CT), and fused PET/CT images (DL_FU). The best-performing radiomics and DL models were combined with dosiomics to create the final joint model. ROC curves with AUC, accuracy, sensitivity, and specificity evaluated the performance. A nomogram was constructed using top-performing model features, parameters, and relevant clinical factors. RESULTS: The extreme gradient boosting (XGBoost) and 18-layer residual neural network (Resnet-18) achieved the best performance. The R+D+DL model combined radiomics, dosiomics, and DL features achieved AUCs of 0.93, 0.92 and 0.89 in the training, internal validaiton and external validation cohorts, respectively. A nomogram constructed with gender, Adaptive RT, SUVp90, and XGBoost-score achieved an AUC of 0.94 for RP prediction in VMAT-treated lung cancer patients using PET/CT. CONCLUSION: Integrating radiomics, DL, dosiomics features and SUVp90 is promising in the RP prediction for lung cancer patients underwent VMAT using PET/CT images.

The design of neural networks typically involves trial-and-error, a time-consuming process for obtaining an optimal architecture, even for experienced researchers. Additionally, it is widely accepted that loss functions of deep neural networks are generally non-convex with respect to the parameters to be optimised. We propose the Layer-wise Convex Theorem to ensure that the loss is convex with respect to the parameters of a given layer, achieved by constraining each layer to be an overdetermined system of non-linear equations. Based on this theorem, we developed an end-to-end algorithm (the AutoNet) to automatically generate layer-wise convex networks (LCNs) for any given training set. We then demonstrate the performance of the AutoNet-generated LCNs (AutoNet-LCNs) compared to state-of-the-art models on three electrocardiogram (ECG) classification benchmark datasets, with further validation on two non-ECG benchmark datasets for more general tasks. The AutoNet-LCN was able to find networks customised for each dataset without manual fine-tuning under 2 GPU-hours, and the resulting networks outperformed the state-of-the-art models with fewer than 5% parameters on all the above five benchmark datasets. The efficiency and robustness of the AutoNet-LCN markedly reduce model discovery costs and enable efficient training of deep learning models in resource-constrained settings.