Explainable handcrafted features for mitotic event detection and classification.

Analysis of cell populations and their behavior is very important in biological and medical research, such as tissue engineering and cancer research. Such behavioral analysis is often performed by visual observation of the cells using microscopy and other imaging techniques, which is often a time-consuming process due to the need for manual observations. Here, a fully automated mitotic event detection method has been developed allowing to reduce the processing time and improving the accuracy of the proliferation rate estimation of studied cell populations Despite the obvious morphological changes during mitosis, traditional image processing methods are not the best candidates, because some phenomena in the images appear as mitotic events but consist of noise and artifacts. These unwanted noise elements and artifacts can cause false positive detections and lead to low precision in detecting proliferation processes. Additionally, traditional cell imaging methods are based on single-cell segmentation resulting in lower performance for high cell densities. To minimize false positive detections, a machine learning-based modeling approach was developed and tested on two large datasets, one public dataset containing phase contrast images and another containing lens-free images. In the new method, traditional image processing, such as thresholding and cell tracking, is employed to select the candidate events. Next, the features that relate to the cell characteristics and process events are extracted. Before the classification step, features are selected using mutual information and ANOVA testing. Finally, false positive rejection is done using tree and random forest classifiers. The applied machine learning approach not only allows high processing performance but also explains how selected features contribute to mitotic event detection. The mean accuracy of the classifiers is 85.12% and precision and recall for the publicly available phase contrast dataset are 88.01% and 92.70% respectively. When applied to the lens-free image dataset, accuracy, precision, and recall were even higher and amounted to 87.66% 88.01%, and 91.78% respectively while employing much less features. The developed methodology has similar performance compared to the published deep learning approach, but has the advantage that it combines high performance with explainable features.