A deep learning workflow for quantification of micronuclei in DNA damage studies in cultured cancer cell lines: A proof of principle investigation

Panchbhai, Anand; Ishanzadeh, Munuse Ceyda; Pankanti, Smarana; Sidali, Ahmed; Solaiman, Nadeeen; Kanagaraj, Radhakrishnan; Murphy, John J.; Surendranath, Kalpana

doi:10.1016/j.cmpb.2023.107447

Cited by 2 publications

(1 citation statement)

References 35 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Here, we present micronuclAI, a novel deep-learning-based pipeline for assessment of CIN through automated quantification of MN from nuclei-stained images. micronuclAI distinguishes itself over previous methods [26][27][28][29][30][31][32] as it 1) can quantify for both MN and NBUDs, 2) requires only nuclear staining, 3) is able to work with 10x to 100x image objectives, 4) can work with any segmentation mask, and most importantly 5) is extensively evaluated in multiple cell lines and thus, ready for use by the community (Fig. 1 and Table 1).…”

Section: Mainmentioning

confidence: 99%

micronuclAI: Automated quantification of micronuclei for assessment of chromosomal instability

Ibarra-Arellano,

Caprio,

Hada

et al. 2024

Preprint

View full text Add to dashboard Cite

Chromosomal instability (CIN) is a hallmark of cancer that drives metastasis, immune evasion and treatment resistance. CIN results from chromosome mis-segregation events during anaphase, as excessive chromatin is packaged in micronuclei (MN), that can be enumerated to quantify CIN. Despite recent advancements in automation through computer vision and machine learning, the assessment of CIN remains a predominantly manual and time-consuming task, thus hampering important work in the field. Here, we present micronuclAI, a novel pipeline for automated and reliable quantification of MN of varying size, morphology and location from DNA-only stained images. In micronucleAI, single-cell crops are extracted from high-resolution microscopy images with the help of segmentation masks, which are then used to train a convolutional neural network (CNN) to output the number of MN associated with each cell. The pipeline was evaluated against manual single-cell level counts by experts and against routinely used MN ratio within the complete image. The classifier was able to achieve a weighted F1 score of 0.937 on the test dataset and the complete pipeline can achieve close to human-level performance on various datasets derived from multiple human and murine cancer cell lines. The pipeline achieved a root-mean-square deviation (RMSE) value of 0.0041, an R2 of 0.87 and a Pearson's correlation of 0.938 on images obtained at 10X magnification. We tested the approach in otherwise isogenic cell lines in which we genetically dialed up or down CIN rates, and also on a publicly available image data set (obtained at 100X) and achieved an RMSE value of 0.0159, an R2 of 0.90, and a Pearson's correlation of 0.951. Given the increasing interest in developing therapies for CIN-driven cancers, this method provides an important, scalable, and rapid approach to quantifying CIN on routinely obtained images. We release a GUI-implementation for easy access and utilization of the pipeline.

show abstract