ESTRO 2024 - Abstract Book

S3012

Physics - Autosegmentation

ESTRO 2024

The DIP and DIE are distinct relevant measures to assess the clinical impact of manual adjustments to autosegmentation. For the head and neck OARs investigated, a mean dose difference of ≥2.0 Gy was found using the DIP and DIE analysis for none of the OARs, although the mean dose differences found with the DIE analysis were generally higher. An NTCP-value difference ≥1.0pp was found only using the DIE analysis for grade II xerostomia when using the unadjusted parotid gland contours. The performed analysis indicates that the dosimetric impact of contouring adjustments is limited, with the impact on the optimised treatment plan smaller compared to the impact on plan evaluation, even though notable geometric adjustments to OARs are made, indicated by the mean DSC ranging from 0.50-0.95. The results show that DLC inaccuracies can be clinically acceptable, which could lead to a more efficient automated workflow for head and neck cancer patients.

Keywords: Dosimetric impact, Autosegmentation, Head and neck

598

Digital Poster

Learning without forgetting improves brain metastases autodetection in multicenter collaboration

Yixing Huang 1 , Zahra Khodabakhshi 2 , Ahmed Gomaa 1 , Rainer Fietkau 1 , Matthias Guckenberger 2 , Christoph Bert 1 , Andratschke Nicolaus 2 , Stephanie Tanadini-Lang 2 , Florian Putz 1 1 University Hospital Erlangen, Department of Radiation Oncology, Erlangen, Germany. 2 University Hospital Zurich, Department of Radiation Oncology, Zurich, Switzerland

Purpose/Objective:

Due to data privacy regulations, medical data sharing among multiple centers is restricted, which hinders the development and validation of deep learning models. Peer-to-peer federated learning (P2PFL) can promote multicenter collaboration by sharing trained models instead of raw data without requiring a complex centralized infrastructure. Continual learning techniques can mitigate the forgetting problem after model sharing. This work aims to investigate the feasibility of continual learning for multicenter collaboration on brain metastases autodetection.

Material/Methods:

T1 contrast enhanced MRI datasets from University Hospital Erlangen, University Hospital Zurich, and BraTS 2023 Brain Metastases Segmentation Challenge were used for evaluation (dataset details see Table 1). The BraTS (training) dataset is publicly available. Feature differences in metastasis spatial distribution and image quality are observed among these centers. In our previous comprehensive survey [1], learning without forgetting (LWF) [2] was demonstrated to have superior performance to other continual learning regularization methods [3]. Therefore, LWF is applied in this work. LWF utilizes the model from the previous center as a teacher model to guide the training of the new (student) model. The knowledge distillation loss (KDL) [4] is used to preserve learned knowledge from the teacher model. In addition, a