ESTRO 2024 - Abstract Book

S3160

Physics - Autosegmentation

ESTRO 2024

[4] F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation,” Nat Methods, vol. 18, no. 2, pp. 203–211, 2021, doi: 10.1038/s41592-020-01008-z.

[5] The MathWorks Inc., “Statistical and machine learning functions for data analysis:12.2(R2021b).” 2021. [Online]. Available: https://uk.mathworks.com/products/matlab.html

3117

Digital Poster

A study of the clinical validity of two commercial AI autocontouring models for radiotherapy.

Daniel Díaz Martín 1 , Artur Sampayo Muñoz 1 , Breixo Carmelo Portas Ferradás 1 , Abel Niebla Piñero 1 , Antonio Salinas Martín 2 , Claudio Fuentes Sánchez 2 , Miriam Vera Dumpiérrez 2 , Leonardo Lorenzo Quintero 2 , Lidia Gómez Perea 2 , Sadia Tremolada Erausquin 2 1 Hospital Universitario Nuestra Señora de la Candelaria, Medical Physics and Radiation Protection Department, Tenerife, Spain. 2 Hospital Universitario Nuestra Señora de la Candelaria, Radiation Oncology Department, Tenerife, Spain

Purpose/Objective:

In radiotherapy, manual contouring is time-consuming and can result in dosimetric variations in patient treatments due to inter- and intra-observer variability. Deep learning-based autocontouring offers a solution to address these challenges. This study evaluates the artificial intelligence (AI) auto-segmentation contours generated by two commercial vendors: Siemens syngo.via RT Image Suite VB40 software and Raystation to assess their clinical usability [ 1,2 ].

Material/Methods:

A total of 39 patients with prostate tumors were randomly and retrospectively selected. All CT scans were performed using Siemens SOMATOM go.Open Pro scanner. Organs at risk (OARs), including the bladder, rectum, and both left and right femoral heads, clinically used for treatment, were auto-contoured using Siemens Syngo.via's AI software and reviewed by technicians and radiation oncologists following internal protocols and international guidelines. For each case, the automatically generated AI contours by Siemens were compared to a second AI segmentation done by the Raystation treatment planning system. Both systems employ deep learning model segmentation that are implemented independently in different anatomical locations. Quantitative evaluation metrics were employed, with a focus on Volumetric Similarity (VS), a measure that assesses the similarity of volumes (closer to 1 indicating more similarity) [ 3 ]. However, these metrics alone may not fully capture clinical utility or expert opinions. Therefore, a qualitative evaluation was conducted, divided into two subsets, involving five expert radiation oncologists and one junior oncologist, using a 3-point Likert scale ( Table 1 ), similar to