ESTRO 2024 - Abstract Book

S4478

Physics - Machine learning models and clinical applications

ESTRO 2024

This study evaluates the performance of an AI model in identifying cancerous patients and prioritizing colonoscopy referrals using a large tabular dataset of patients in Québec, Canada. Despite the significant imbalance data and the presence of many missing values for various features, our AI-based solution achieved high performance in predicting cancer, with promising results in the prioritization task, while utilizing readily available, limited clinical information.

Keywords: triage, crc prediction, colonoscopy referral

References:

[1] Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: synthetic minority over-sampling technique. Journal of artificial intelligence research, 16, 321-357.

[2] Ayling, R. M., Wong, A., & Cotter, F. (2021). Use of ColonFlag score for prioritisation of endoscopy in colorectal cancer. BMJ Open Gastroenterology, 8(1), e000639.

[3] Ayling, R. M., Lewis, S. J., & Cotter, F. (2019). Potential roles of artificial intelligence learning and faecal immunochemical testing for prioritisation of colonoscopy in anaemia. British Journal of Haematology, 185(2), 311 316.

1455

Proffered Paper

Deep-learning deformable image registration for 4DCT/3DMRI label prop-agation.

Ziad Kheil 1,2 , Soleakhena Ken 1,2 , Laurent Risser 3,4

1 Institut Claudius Regaud/Institut Universitaire du Cancer de Toulouse – Oncopôle, Engineering and Medical Physics, Toulouse, France. 2 Inserm U1037- Centre de Recherches Contre le Cancer de Toulouse, Radiation Oncology, Toulouse, France. 3 CNRS, Université de Toulouse, Toulouse, France. 4 Institut de Mathématiques de Toulouse, Université de Toulouse, Toulouse, France

Purpose/Objective:

Stereotactic body radiation therapy (SBRT) delivers high gradient dose and requires accurate tumour and organs at risk segmentation. In the case of mobile tumours, segmentation needs to be performed for each respiratory phase of the patient and 4DCT might encounter difficulties to depict highly deformable movements due to breathing motion. Advances in Deep Learning based Deformable Image Registration offer a reliable, and time-efficient method to infer deformation fields mapping an image to another. We propose a deep learning approach in order to warp annotations from a 3DMRI to all phases of a 4DCT using automatically generated smooth, non-rigid deformation fields. Different parts of the pipeline have been validated independently on robust public datasets. Furthermore, the complete pipeline is currently being tested on an in-house dataset of patients to transfer accurate tumour annotations onto the poorly/not visible slices of the 4DCT scans.