RadGuide-S3P-Net: A radiomics-guided self-training semi-supervised deep learning framework for multi-class pneumonia subtype classification

To reduce the dependency on large-scale labeled data in medical imaging, this study proposes RadGuide-S3P-Net, a radiomics-guided self-training semi-supervised neural network for pneumonia subtype classification. The core idea of RadGuide-S3P-Net is to use radiomics-based classical machine learning models trained on a small number of annotated CT images to guide a deep neural network via soft-label distillation. By transferring radiomics-derived priors to the deep model, unlabeled CT scans can be effectively utilized, thus enabling knowledge distillation from a data-efficient “teacher” model to a data-hungry “student” network. The pipeline of the RadGuide-S3P-Net consists of four major stages: segmentation, radiomics feature extraction, soft-label generation, and radiomics-supervised deep learning. A total of 1148 chest CT scans detected with pneumonia by a commercial pneumonia diagnosis and segmentation network were retrospectively included. Overall, four types of pneumonia including bacterial pneumonia, COVID-19, invasive pulmonary aspergillosis, and pulmonary tuberculosis, were identified. Then, the image data were randomly divided into training (70%, n = 803) and test sets (30%, n = 345). With only 30% of training cases labeled and 70% masked, the RadGuide-S3P-Net achieved a test macro-AUC of 0.9170, outperforming the radiomics model (macro-AUC: 0.8853, P = 0.0275) and a 3D‑ResNet‑18 trained with 30% labels (macro-AUC: 0.8495, P = 0.0001), and achieving performance comparable to a fully supervised 3D‑ResNet‑18 trained on all annotations (macro-AUC: 0.9178, P = 0.4924). These results demonstrate that radiomics-guided soft-label distillation can effectively exploit unlabeled medical images to train high-performing neural networks for multi-class pneumonia diagnosis.