ESTRO 2025 - Abstract Book

S1370

Clinical - Lung

ESTRO 2025

Purpose/Objective: Chemoradiation (CRT) is the cornerstone for the treatment of locally advanced stage III non small cell lung cancer (NSCLC). Multimodal tools can be useful to predict prognosis after CRT and therefore allow to adopt descalating or escalating approaches in order to reduce additional unnecessary toxicity or improve outcomes when needed Material/Methods: This study used images and clinical data from the CLARO dataset, which includes stage III locally advanced NSCLC patients treated with concurrent chemoradiation. Baseline simulation CT scans were used as images. Clinical data included in the model were tumor stage, T and N staging, age, sex, histology, and initial CTV, for a total of seven descriptors. To predict prognosis we utilized a multimodal deep learning pipeline. This pipeline integrates models from different modalities using multiobjective optimization. By leveraging diverse AI models, we selected the most effective ones to form an ensemble. We trained multiple unimodal models on clinical data and imaging data. For clinical data, we used seven machine learning models. For imaging data, we utilized 30 pretrained convolutional neural networks from eight architecture families. The optimal ensemble was selected through multiobjective optimization, balancing recall and kappa diversity to ensure high sensitivity and complementary model errors Results: Clinical data ed images of 160 stage III NSCLC patients were included in the model.The majority of patients were male (56.25%) and the most common histology was adenocarcinoma (63.75%) (see table 1 for clinical data included in the model).