Author(s)

Fenwei Guo

Corresponding Author

Fenwei Guo

ABSTRACT

Delineating precisely the locations of individual neuronal cells in microscopic images is of great significance for the treatment of neurological disorders and neurodegenerative diseases. However, conventional methods for the instance segmentation of neuronal cells suffer from the limited accuracy, lack of automation, and time intensive processes. To address this challenge, we propose an R-CNN-based deep learning model for the segmentation of neuronal cells with a promising performance in this paper. The architecture of our model is the Cascade Mask R-CNN, which is a combination of the Mask R-CNN and Cascade R-CNN. In this model, a ResNeXt + FPN backbone with standard convolution and fully connected heads is utilized for the mask prediction, where the ResNeXt part of backbone is ResNeXt-152-32x8d. The model is pretrained based on the LIVEcell dataset, and subsequently trained using the dataset provided by Sartorius in a Kaggle competition. By a boost from the pseudo-label technique, our model can achieve a [email protected]:.95 score 0.338 on the private test set. Such a score locates at 36/1505 (top 3%) in the leaderboard of Sartorius - Cell Instance Segmentation competition, and can get a silver medal in this Kaggle competition. Our results could help the researchers measure the effects of neurological disorders more easily, and potentially accelerate the discovery and development of new drugs for the treatment of neurodegenerative diseases.

KEYWORDS

Instance segmentation, Neuronal cell, deep learning, Mask R-CNN, Cascade R-CNN