Feasibility study on automated cytokinesis-block micronucleus assay analysis in cytogenetic biodosimetry using YOLOv5 object detection.

PURPOSE: Accurate dose estimation is crucial in radiation emergency medicine to predict potential clinical outcomes and to develop appropriate treatment plans. This need becomes especially important during mass-casualty events, where reliable and rapid triage is necessary. However, cytogenetic biodosimetry, which can be used for triage, is bottlenecked by the time required for cell culture and the expertise needed of chromosomal analysis. The objective of this study is to apply deep learning-based object detection to the analysis of micronuclei (MNs).

MATERIALS AND

METHODS

Peripheral blood samples were collected from healthy volunteers with informed consent. For model training and validation, samples were irradiated at 0 (sham), 2, and 3 Gy. Whole blood cultures were stimulated with phytohemagglutinin and treated with cytochalasin B (at 44 h) for 72 h. Cells were scanned for whole slide imaging. M1-M4 cells were annotated for nuclear division index (NDI) analysis, and main nuclei and MNs in binucleated cells (M2) were annotated for MNs analysis. Both the NDI and MNs models were trained using the YOLOv5 framework. Dose-response curves generated by the deep learning-based models were compared with previously published manually scored curves.

RESULTS

Although still in the preliminary stages, we confirmed that deep learning-based object detection using YOLOv5 can achieve good classification performance. There is a possibility for further improvement of the model using data augmentation, particularly for a low number of training images. The dose-response curves derived from deep learning-based analysis were comparable to previously reported manual calibration curves.

CONCLUSION

The use of deep learning techniques for image recognition offers a promising approach for rapid and reliable NDI and MNs detection in cytogenetic biodosimetry.