Alessia De Biase

Purpose Within the medical field, there is a growing demand for deep learning (DL) models which convey model certainty to the end-user, while maintaining alignment between model accuracy and certainty. For oropharyngeal cancer (OPC) primary tumour (PT) segmentation in PET/CT images, an ensemble-based DL model was developed which outputs tumour probability maps (TPM) showing voxel-level predicted probabilities. This study extended the network to generate TPMs for both PT and pathologic lymph nodes (PL) and explored whether quantified uncertainty in TPMs can predict segmentation model accuracy in an independent external cohort. Methods We gathered PET/CT images and manual delineations of gross tumour volume of the PT (GTVp) and PL (GTVln) of 405 OPC patients treated with (chemo)radiation in our institute in 2010-2022. The publicly available 2022 HECKTOR challenge dataset served as external test set. The existing network was adapted to perform multi-label segmentation, training 15 models and considering the ensemble average of predicted TPMs per patient. Surface and aggregate DSC were computed for the predicted contours at different probability thresholds. Uncertainty was quantified by coefficient of variation (CV) of multiple predictions. Results GTVln segmentation showed slightly lower performance than GTVp: aggregate DSC of 0.66 and 0.67 in the internal, 0.70 and 0.75 in the external test sets. However, a significant negative correlation (about -0.6) was observed in both test sets between Average Surface DSC and CV for both structures, indicating a significant calibration. Conclusion Significant accuracy versus uncertainty calibration was achieved for TPMs in both internal and external test sets, indicating the potential use of CV to identify cases with lower GTVp and GTVln segmentation accuracy, independently of the dataset.

Radiation therapy is a localized cancer treatment that relies on precise delineation of the target to be treated and healthy tissues to guarantee optimal treatment effect. This step, known as contouring or segmentation, involves identifying both target volumes and organs at risk on imaging modalities like CT, PET, and MRI to guide radiation delivery. Manual segmentation, however, is time-consuming and highly subjective, despite the presence of contouring guidelines. In recent years, automated segmentation methods, particularly deep learning models, have shown promise in addressing this task. However, challenges persist in their clinical use, including the need for robust quality assurance (QA) processes and addressing clinical risks associated with the use of the models. This review examines the challenges and considerations of the clinical adoption of deep learning target auto-segmentation in radiotherapy, focused on the target volume. We discuss potential clinical risks (eg, over- and under-segmentation, automation bias, and appropriate trust), mitigation strategies (eg, human oversight, uncertainty quantification, and education of clinical professionals), and we highlight the importance of expanding QA to include geometric, dose-volume, and outcome-based performance monitoring. While deep learning target auto-segmentation offers significant potential benefits, careful attention to clinical risks and rigorous QA measures are essential for its successful integration in clinical practice.

**Abstract text** Within the medical field, there is a growing demand for deploying deep learning (DL) models which convey model certainty to the | POSTER

**Abstract text** Stereotactic body radiotherapy (SBRT) is the standard approach for patients with early stage lung cancer who are inoperable bec | POSTER