Brain Cell Segmentation from the LIVECell Dataset Using Deep Learning

Alzheimer's disease and brain tumors are two of the world's top causes of death and disability. It is challenging to assess the efficacy of treatments for potentially deadly conditions. Neuron segmentation is complex and labor-intensive. Due to their shifting shape and cell stacking, neuronal precision segmentation has diminished. Precise image segmentation facilitates the study of complex biological themes, yet high-density, low-contrast objects may need more elaborate image processing pipelines. Despite being the most cutting-edge method, there is a crucial need for deep learning-based image segmentation models. This study segments the label-free LIVECell-large-scale dataset. The dataset consists of 5239 carefully annotated and expert-validated records encoded using the Run-length Encoding (RLE) technique. These cell lines were selected to expand LIVECell's machine learning capabilities. Three segmentation models are used in this study: EfficientDet, U-net, and Faster R-CNN. Two experiments are used to train the model. LIVECell training data is employed in cell classes to analyze the segment performance of each class. The LIVECell dataset was trained with EfficientDet, U-net, and the Faster R-CNN segmentation model. As a result, the Intersection over union (IoU) value is between 50.0% and 100.0%, with Average Precision (AP) values of 58.824%, 42.655%, and 51.625%, respectively; Average False-Negative Ratio (AFNR) values are 56.566%, 36.634% and 48.745%, respectively. Training dataset scaling tests have shown that AP value increases with dataset size when training is performed at the same IoU threshold value.