Convolutional neural networks (CNNs) have emerged as powerful tools in medical imaging for predicting cancer outcomes. While traditional approaches use CNNs for binary classification—such as determining whether a patient will develop distant metastasis (DM)—this study extends their application to time-to-event analysis. By incorporating censoring information, the proposed 2D and 3D CNN models generate DM-free probability curves over time for individual patients, enabling more nuanced prognostic assessments. The model was trained and validated on 294 patients from a benchmark cohort, then tested on three independent cohorts totaling 743 patients: MAASTRO (n=136), PMH (n=497), and CRO (n=110). A total of 1,037 patients were included across all datasets.
The best-performing model achieved Harrell’s concordance indices (HCIs) of 0.88, 0.67, and 0.77 in the three testing cohorts, with two out of three showing strong performance comparable to the 3-fold cross-validation results (HCIs of 0.78, 0.74, and 0.80). This indicates that the model generalizes well across diverse clinical settings. Additionally, the CNNs demonstrated significant ability to stratify patients into high- and low-risk groups, confirmed by log-rank tests (p < 0.05) across all testing sets. The 2D-CNN and 3D-CNN+Clinical models showed particularly robust stratification, suggesting their potential utility in personalized treatment planning. To assess the role of image texture, a binary masking experiment was conducted, replacing Hounsfield Unit values within the gross tumor volume (GTV) with uniform +1 and -1 values. The resulting performance drop was minimal, especially for the 3D-CNN, indicating that volumetric shape and size are dominant predictors.NMNAT1 Antibody Description This supports the idea that deep learning models can extract meaningful prognostic signals from basic anatomical features without relying heavily on complex texture patterns.
Compared to an artificial neural network (ANN) based on clinical variables alone, the CNNs outperformed or matched its performance, particularly in the MAASTRO and CRO cohorts.COX IV Antibody supplier However, combining clinical data with imaging inputs (CNN+Clinical) did not significantly improve results, possibly due to limited clinical variable availability across datasets.PMID:35085381 The study highlights the reliability of CNN-based models in DM prediction and underscores the value of integrating time-to-event modeling in radiomics. These findings support the development of precision radiation oncology strategies, where imaging-driven deep learning can guide dose adaptation and risk mitigation.MedChemExpress (MCE) offers a wide range of high-quality research chemicals and biochemicals (novel life-science reagents, reference compounds and natural compounds) for scientific use. We have professionally experienced and friendly staff to meet your needs. We are a competent and trustworthy partner for your research and scientific projects.Related websites: https://www.medchemexpress.com