Abstract:
This study investigates the performance of 28 different UNet models for
segmenting and determining cell confluence in brightfield microscopy images,
combining various hyperparameters such as loss functions, batch sizes, and epochs.
Ground truths for the images were manually annotated which was another challenge
of this study. Among the models, two of them were chosen since they achieved high
accuracy results. The study also evaluates the effects of different biomaterial density
on cell growth using these models. The results showed that low-density biomaterials
(5 ug) were non-toxic, while medium (20 ug) and high concentrations (50 ug for
PAR30 and 500 ug for PLL250) significantly suppress cell growth, with confluence
ratios dropping below 70%. Additionally, various classification models were tested on
datasets with different cell images and biomaterial densities. Principal Component
Analysis (PCA) and hybrid models were found to significantly improve classification
accuracy, particularly in binary classification tasks, which achieved accuracies nearing
98%. The study highlights the performance of different model architectures, manual
annotation for ground truth, and dimensionality reduction techniques in enhancing the
accuracy of cell confluence segmentation and biomaterial toxicity assessment.
