S2493
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
fold cross-validation approach. Thek-fold model with minimum validation loss was chosen, and evaluated by means of AUC-ROC, precision, recall and balanced accuracy. Thismodel was then externally testedto infer benign/malignant predictions on theremaining patients of the SBRT cohort (294 in total). To evaluate performance in this subgroup, Kaplan-Meier curves were generated for each benign/malignant groups to contrast the validity of our predictions. Results: The selected model presented a performance AUC- ROC of 0.87, precision of 0.47, recall of 0.68 and balanced accuracy of 0.77 (Figure1). After using the model to infer predictions on SBRT patients, 80% (93/116) and 20% (23/116) of SBRT patientswtPC, and84%(149/178) and 16% (29/178) of SBRT patientswoPC, got a malignant and benign prediction, respectively. Kaplan-Meyer curves fit on each predicted subgroup matched the expected overall survival in each case (Figure2).
Digital Poster Highlight 4744
Malignancy prediction through CT foundation model for lung lesions without pathologic confirmation Alba Magallon-Baro 1,2 , Ditsuhi Iskandaryan 3,2 , Ruben Moreno-Aguado 1,2 , Samantha Aso 4 , Susana Padrones 4 , Anna Ureña 5,6 , Victor Moreno 1,2 , Arturo Navarro- Martin 7,8 1 Oncology Data Analytics Program, Catalan Institute of Oncology, Bacelona, Spain. 2 Colorectal Cancer Group ONCOBELL Program, nstitut de Recerca Biomedica de Bellvitge (IDIBELL), Barcelona, Spain. 3 Oncology Data Analytics Program, Catalan Institute of Oncology, Barcelona, Spain. 4 Pulmonary Department, Hospital Universitario Bellvitge, Barceloa, Spain. 5 Thoracic Surgery Department, Hospital Clínic, Barcelona, Spain. 6 Inflammation and Repair in Respiratory diseases, Institut d’Investigacions August Pi i Sunyer (IDIBAPS), Barcelona, Spain. 7 Radiation Oncology Department, Hospital Clínic, Barcelona, Spain. 8 Translational Genomics and Targeted Therapies in Solid Tumors Group, Institut d’Investigacions August Pi i Sunyer (IDIBAPS), Barcelona, Spain
Purpose/Objective: Non-resectable patients with solitary lung
nodules receive stereotactic radiotherapy (SBRT) when suspected of primary lung cancer (LC)at early stage. An important fraction of these patients receives SBRT without pathologic confirmation (PC) of malignancy due to fragility or other risks associated with biopsy, which may result into a small fraction of patients receiving treatment despite having a benign tumour.The aim of this study was to developan artificial intelligence (AI)-based modelusing patients undergoing both surgery or SBRT with confirmedPC (w tPC),to predictradiological features in diagnostic CT scans and use themto identifybenign/malignant lesionson SBRT patients without PC (woPC). Material/Methods: A total of 1048 diagnostic CT scans were retrospectively collected for 572 and 410LC patients undergoing surgery or SBRT, respectively.The surgical cohort: 462 patients presented malignant tumourand 110 benign tumour.The SBRT cohort, we 232wtPCand 178woPC patients.VoxelFM- acustom 3D foundation model(FM) for CT scans based on a modified DINOv2 architecture- was used to extract radiological featureson all diagnostic CTs. Next,features of all surgical (462) and half SBRT (116) patients wtPC(6 88 in total) were used for training and validatinga perceiver transformer-based model to predict tumour benign/malignancy, following a 5-
Conclusion: Diagnostic CTs showed predictive value to identify benign/malignant tumours. Novel VoxelFMfoundation model
Made with FlippingBook - Share PDF online