An End-to-End Chromosome-Image Screening Framework for Radiation Biodosimetry

Taeyeong Bak, Wol Soon Jo, Soo Kyung Jeong, Su Jung Oh, Jimin Lee

Keywords: biodosimetry, chromosome aberration, dicentric chromosome, translocation, object detection

Registration Requirement: Yes

Abstract: Radiation biodosimetry relies on chromosome-aberration analysis in blood lymphocytes, but manual scoring of dicentric chromosomes and translocations remains slow and expert-dependent. Prior studies have demonstrated automated dicentric-based biodosimetry, while translocation or FISH-based biodosimetry remains important for retrospective assessment. Deep-learning-based cytogenetic image analysis has also been demonstrated in fluorescence mFISH segmentation, and modern object detection has been applied to chromosomal-aberration analysis in a biodosimetry-adjacent setting. However, integrated modern image-based joint screening across both axes remains limited. We present an end-to-end chromosome-image screening framework that automates this pre-screening step with RT-DETR. Using a 15,450-image subset from a total of 18,818 images, we simplify the label space to translocation(tr), dicentric chromosome(dic), and chromosome(chr). On the test split, AP50 reaches 0.758 for dic and 0.738 for tr. These results indicate meaningful localization of abnormality candidates in real data, although severe class imbalance, simplified labels, and scorable single-cell identification remain major bottlenecks.

Read CFP & Author Instructions: Yes

Originality Policy: Yes

Single-blind & Not Under Review Elsewhere: Yes

LLM Policy: Yes

Submission Number: 49