Speaker
Description
For the detection of neuronal cell bodies in 1-micron BigBrain data we propose a conceptually simple framework called Contour Proposal Network (CPN). The CPN detects and segments possibly overlapping cells by fitting closed contours using a fixed-sized representation based on Fourier Descriptors. State-of-the-art object detection architectures can be used as backbone networks, forming a single-stage instance segmentation model that is trained end-to-end. We evaluate the CPN with different backbone networks using datasets from different modalities, including the 1-micron BigBrain. Experiments show that CPNs outperform U-Net and Mask R-CNN in instance segmentation accuracy. The CPN is computationally very efficient and is suitable for real-time applications when coupled with backbones such as ResNet-50 FPN. The trained models generalize well, even across different domains of cell types. The main assumption of the method regards closed object contours, hence the CPN is applicable to a wide range of detection problems also outside the biomedical domain. PyTorch code has been made available at: celldetection.org
