1127 - Prediction of Radiotherapy Efficacy of Esophageal Cancer Using Radiomics Temporal Sequence Extracted from Weekly CBCT

H. Wu¹, T. Tang², R. Mao³, and J. Zhang⁴; ¹The Affiliated Cancer Hospital of Zhengzhou University, ZhengZhou, China, ²General hospital of pingmei shenma group, PingdingShan, Henan, China, ³The Affiliated Cancer Hospital of Zhengzhou University, Zhengzhou, China, ⁴Mount Sinai South Nassau, Oceanside, NY

Purpose/Objective(s): The aim of this study was to assess the predictive value of radiomics temporal sequences in indicating treatment response and to develop a rapid and reliable method for predicting treatment efficacy. Materials/

Methods: A cohort of 121 patients diagnosed with stage II–IVa esophageal squamous cell carcinoma (ESCC) who underwent intensity-modulated radiation therapy (IMRT) was selected. Nineteen radiomics features were extracted from target and organs at risk (OAR), generating five sets of data for each patient from weekly cone-beam computed tomography (CBCT) scans using GraphPad Prism 8. Patients were categorized into three groups based on survival time (short, median, long). A long short-term memory (LSTM) sequence classifier was developed using a custom Matlab script to predict treatment efficacy post-treatment, while another LSTM model was trained to forecast radiomics response for future treatment fractions. Bootstrap technique was employed to assess model confidence.
Results: The LSTM classifier and sequence predictor were trained on 70% of the patient cohort and validated on the remaining 30%. The classifier exhibited an average accuracy of 91% (±6.7%) when utilizing radiomic time sequences. The sequence predictor demonstrated an average accuracy of 86% (±5.3%) for the last two weeks of radiomics features during treatment. When applied to the entire predicted sequence, the classifier achieved an accuracy of 79% (±11%).
Conclusion: Two LSTM models, a classifier, and a sequence predictor were developed with notable accuracy based on patients radiomics sequences. The predictor can anticipate radiation response before treatment completion, facilitating evaluation of efficacy and guiding future clinical interventions. Future efforts will focus on enhancing the robustness of the models.