ESTRO 2025 - Abstract Book

S2435

Physics - Autosegmentation

ESTRO 2025

GORTEC, HKNPCSG, HNCIG, IAG-KHT, LPRHHT, NCIC CTG, NCRI, NRG Oncology, PHNS, SBRT, SOMERA, SRO, SSHNO, TROG consensus guidelines. Radiother Oncol . 2018;126(1):3-24. doi:10.1016/j.radonc.2017.10.016

1450

Digital Poster Artificial Intelligence-based Cell Survival Colony Counting Model Wanrong Mona Wang 1 , Joanna Li 2,3 , Laya Rafiee Sevyeri 2,3 , Shirin Abbasi Nejad Enger 2,3

1 Department of Bioengineering, Faculty of Engineering, McGill University, Montréal, Canada. 2 Medical Physics Unit, Department of Oncology, Faculty of Medicine, McGill University, Montréal, Canada. 3 Lady Davis Institute for Medical Research, Jewish General Hospital, Montréal, Canada Purpose/Objective: Clonogenic assay is the gold-standard technique in radiobiology for assessing the radiosensitivity of cancer cells and for quantifying the effectiveness of radiotherapy; it does so by manually counting 1 colonies that survive post irradiation, where a viable colony is considered to contain more than 50 cells 2 . In this study, we propose an artificial intelligence (AI)-based solution to automate the clonogenic assay counting process. Material/Methods: We collected 787 digital images of clonogenic assays from HCT116 human colorectal cancer cell line and developed a pipeline to automate colony counting to address several challenges: (i) the lack of annotated data due to the laborious nature of manually labeling colonies, (ii) variability in data collection, which leads to visible differences in images that require preprocessing, (iii) the absence of segmentation masks for supervised deep learning (DL) models due to the lack of annotations, and (iv) the unavailability of microscopic-resolution images, which necessitates the use of morphological features for accurate counting. Our proposed pipeline utilizes OpenCV 3 to generate a fully segmented clonogenic assay and to predict viable colony counts. Our solution focuses on: (i) the ability to distinguish colonies from non-colonies, attained by size- and color based threshold, (ii) the ability to distinguish bordering colonies (ii) the ability to filter the background beyond the boundary of the well, besides generating segmentation masks suitable for supervised DL models. We first preprocessed the images (Figure 1) to address the presence of background objects and the “edge issue”—a common challenge in distinguishing cells growing along the periphery of the well from the edge based on colorimetric differences—we employed the HoughCircles function 4 to detect circular wells and generate a binary mask to exclude background objects. The resulting images were then cropped and resized to standardize the dimensions.