3751 - Multidimensional Feature-Based Deep Learning Fusion Model for Non-Invasive Prediction of Occult Lymph Node Metastasis in Early Lung Adenocarcinoma
Presenter(s)
X. Yin1, Y. Lu2, Y. Cui1, J. Wen3, J. Yu4, and X. Meng5; 1Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, 2School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China, 3Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences Department of Radiation Oncology, Jinan, Shandong, China, 4Department of Radiation Oncology and Shandong Provincial Key Laboratory of Precision Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, 5Cheeloo College of Medicine, Shandong University, Jinan, Shandong, China
Purpose/Objective(s): Surgery and stereotactic body radiotherapy (SBRT) are promising curative treatments for early non-small cell lung cancer (NSCLC). However, ignoring the presence of occult lymph node metastasis (OLNM) poses a significant threat to treatment outcomes and significantly increases the risk of recurrence. Thus, our goal is to develop and validate a CT based radiomics and deep learning (DL) fusion model for non-invasive prediction of OLNM.
Materials/Methods: Patients at two centers diagnosed with lung invasive adenocarcinoma with radiological node-negative were included. We developed clinical, radiomics, and combined models using logistic regression. A DL model was established based on the 3D SE-ResNet34 framework, and subsequently, a DL fusion model was created by integrating clinical, radiomics features and deep learning features. The performance of the models was evaluated using the area under the curve (AUC) of the receiver operating characteristic curve, calibration curve and decision curve analysis. Five predictive models were compared to identify the optimal classification model, and SHAP and Grad-CAM methods were used for model visualization and interpretation.
Results: A total of 358 patients were retrospectively enrolled, with 186 allocated to the training cohort, 48 to the internal validation cohort and 124 cases to the external testing cohort. Age, gender, largest tumor diameter, nodule density, and vascular bundle sign were selected as independent predictive factors for OLNM. The DL fusion model was developed by integrating DL features extracted from the 3D SE ResNet-34 network and selected clinical and radiomics features. The model achieved the highest AUC of 0.947 in the training set. It performed well in both internal and external cohorts, reaching AUCs of 0.903 and 0.907, respectively, and significantly better than single-modal DL models, clinical model, radiomics model and radiomics-clinical combined model (Delong test: P<0.05). To better understand the features captured by the predictive model, we identified the regions with the greatest contribution to predictions, highlighting the significance of texture features and the surrounding tumor region in the stratification of OLNM patients.
Conclusion: We have developed a reliable and generalizing DL fusion model that integrates multidimensional features. This model can non-invasively and accurately predict OLNM in early-stage lung adenocarcinoma. For inoperable patients receiving SBRT, since this treatment does not involve prophylactic lymph node irradiation, the model can guide the adjuvant treatment for high-risk OLNM patients after the treatment through OLNM risk stratification. Meanwhile, it can also help operable patients determine the true staging and surgical approaches, thus facilitating the realization of personalized treatment plans.