Author: Xiaolong Fu, Runping Hou, Md Tauhidul Islam, Lei Xing 👨🔬
Affiliation: Department of Radiation Oncology, Stanford University, Department of Radiation Oncology, Shanghai Chest Hospital, Shanghai Jiao Tong University School of Medicine 🌍
Purpose: To introduce a novel schematic image representation of radiomics data, called OmicsMap, for high-performance deep radiomics analysis. OmicsMap transforms tabular radiomics data into an image format where relationships among the features are encoded in its spatial configuration. We demonstrate that integrating OmicsMap with convolutional neural networks (CNNs) leads to substantially improved prognostic predictions for patients with locally advanced non-small cell lung cancer (LA-NSCLC).
Methods: The OmicsMap framework consists of three steps. First, high-dimensional (HD) radiomics features are extracted from region of interests in medical images, followed by the selection of discriminative features. Subsequently, these features are transformed into structured 2D OmicsMaps with data-specific feature interactions encoded in the pixelated configuration of the image maps. Finally, deep-level features are extracted from the OmicsMaps using a CNN for prognostic prediction. The framework was validated on 329 LA-NSCLC patients who underwent lung radiation therapy. The prognostic performance of the OmicsMap-based framework is compared to conventional methods, including Cox regression models incorporating either clinical factors or tabular radiomics features. Model performance was assessed using Harrell’s concordance index (C-index) and the time-dependent area under the receiver operating characteristic curve (AUC).
Results: The OmicsMap-based model demonstrated effective prediction of progression-free survival (PFS) in LA-NSCLC patients, with a C-index of 0.70 (95% CI: 0.64-0.75) in the independent testing set. The time-dependent AUCs of the model at 1, 2 and 3 years in the testing set were 0.76, 0.78 and 0.76, respectively, significantly outperforming the conventional Clinical model (AUC: 0.57, 0.57, 0.64; p<0.05). Compared to the traditional radiomics method, our OmicsMap framework significantly improved predictive performance by 7.69%.
Conclusion: The OmicsMap provides an innovative representation of radiomics features and their interrelationships, enabling superior prognostic prediction in LA-NSCLC patients. This deep radiomics approach allows maximal utilization of radiomics data and holds potential for clinical applications in precision oncology.