A Multi-Modal Deep Learning-Based Decision Support System for Individualized Radiotherapy of Non-Small Cell Lung Cancer

Abstract

<h3>Purpose/Objective(s)</h3> In a complex disease such as cancer, interactions between the tumor and host can exist at the molecular, cellular, tissue, and organism levels. The disease and its evolution may be present in multiple modalities across scales such as clinical, genomic, molecular, pathological, and radiological imaging, which poses a challenge to clinicians in making treatment decisions. This is especially true in cancer radiotherapy, where troves of routinely generated data include clinical data (e.g., diagnosis and prognosis), radiotherapy data (e.g., treatment plans and delivered doses), and the anatomical and functional image data (e.g., CT, MRI, PET). AI can be a powerful tool to explore the many potential care paths in the radiotherapeutic management of cancers and to generate the best treatment option specific to each individual patient while maximizing benefits and minimizing toxicities. In this work, we aim to develop a multi-modal AI-empowered clinical decision support system for personalized radiotherapy of non-small cell lung cancer patients. <h3>Materials/Methods</h3> With IRB approval, data of 217 non-small cell lung cancer patients was obtained, which included treatment planning data, images, and clinical notes. Factors extracted from treatment planning system were focused on clinical, demographic, and dose-volume indexes. Radiomics features were extracted using open-source software. The clinical notes were analyzed using natural language processing to cluster phrases into predictive features. The toxicities of each patient were ranked on a RTOG scale of 0-5 on radiation pneumonitis severity, while the outcomes were ranked on a scale of 0-3 based on tumor reduction over 3 years per RECIST 1.1 guidelines. The consolidated and normalized data was input into a deep neural network with five hidden layers. Using 5-fold cross validation testing (172 training patients and 43 testing patients x 5), an optimal set of weights was determined. <h3>Results</h3> The accuracy of the model was 89.4% for predicting whether radiation pneumonitis was severe (RTOG >=2) and was 71.0% in predicting whether the treatment was successful after 3 years. This approach was superior to traditional approaches focusing on a single datatype. The model performed well with up to 5% of the training data missing, reflecting real clinical challenges. <h3>Conclusion</h3> In this work, we've developed a multi-modality-based decision support system for radiation therapy and demonstrated its efficacy in predicting radiation-induced pneumonitis and patient outcome. Future work will be on automating the curation of structured and unstructured patient data to achieve adaptive personalized radiotherapy for more robust and comprehensive clinical decision support in the clinic.