ESTRO 2025 - Abstract Book

S2437

Physics - Autosegmentation

ESTRO 2025

Results: To evaluate the segmentation module, we calculated average accuracy on all the 787 samples, referring to the difference between the number of the colonies counted by the OpenCV moments 3 to the number of colonies in the ground truth (manually counted by lab technician). The model achieved an accuracy of 77%.

Conclusion: We presented an AI-based solution to automatically count the viable colonies in clonogenic assays.

Keywords: Cell counting, Clonogenic assay, AI

References: [1] Zhang, L. (2022). Machine learning for enumeration of cell colony forming units. Visual Computing for Industry, Biomedicine, and Art , 5(1), 26. [2] Franken, N. A., Rodermond, H. M., Stap, J., Haveman, J., & Van Bree, C. (2006). Clonogenic assay of cells in vitro. Nature protocols, 1(5), 2315-2319. [3] Bradski, G. (2000). The OpenCV library. Dr. Dobb’s Journal of Software Tools . [4] Ballard, D. H. (1981). Generalizing the Hough transform to detect arbitrary shapes. Pattern recognition , 13(2), 111 122.

1572

Digital Poster Trainable Morphological Post-Processing Framework for Enhanced Thoracic Organs-at-Risk Segmentation in Radiation Therapy Planning Tzu-Fang Chang 1 , Chao-Chia Lin 2 , Shanq-Jang Ruan 2 , Ting-Feng Hsieh 2 , Yu Jen Wang 1 1 Department of Radiation Oncology, Fu Jen Catholic University Hospital, New Taipei City, Taiwan. 2 Department of Electrical and Computer Engineering, National Taiwan University of Science and Technology, Taipei, Taiwan Purpose/Objective: Precise segmentation of thoracic organs-at-risk (OAR) is critical for optimizing radiation therapy planning, minimizing radiation exposure to healthy tissues, and improving treatment outcomes. Despite significant advancements in deep learning-based automated segmentation, achieving clinical-grade accuracy remains challenging due to anatomical variability and imaging complexities. This study introduces a novel trainable morphological post-processing framework designed to refine and enhance the performance of existing deep learning models for thoracic OAR segmentation. Material/Methods: We developed a framework that integrates trainable dilation and erosion layers equipped with differentiable structuring elements (SEs), analogous to convolutional kernels. These layers perform max-sum and min-sum operations on input feature maps, enabling data-driven refinement of segmentation boundaries and contours. The method is architecture-agnostic and can be seamlessly incorporated into any deep learning model employing pixel wise classification, offering an adaptable tool for automatic segmentation enhancement.