No abstract

No abstract

Connectomics is a developing field aiming at reconstructing the connection of the neural system at the nanometer scale. Computer vision technology, especially deep learning methods used in image processing, has promoted connectomic data analysis to a new era. However, the performance of the state-of-the-art (SOTA) methods still falls behind the demand of scientific research. Inspired by the success of ImageNet, we present an annotated ultra-high resolution image segmentation dataset for cell membrane (U-RISC), which is the largest cell membrane-annotated electron microscopy (EM) dataset with a resolution of 2.18 nm/pixel. Multiple iterative annotations ensured the quality of the dataset. Through an open competition, we reveal that the performance of current deep learning methods still has a considerable gap from the human level, different from ISBI 2012, on which the performance of deep learning is closer to the human level. To explore the causes of this discrepancy, we analyze the neural networks with a visualization method, which is an attribution analysis. We find that the U-RISC requires a larger area around a pixel to predict whether the pixel belongs to the cell membrane or not. Finally, we integrate the currently available methods to provide a new benchmark (0.67, 10% higher than the leader of the competition, 0.61) for cell membrane segmentation on the U-RISC and propose some suggestions in developing deep learning algorithms. The U-RISC dataset and the deep learning codes used in this study are publicly available.

Connectomics is a developing filed aiming at reconstructing the connection of neural system at nanometer scale. Computer vision technology, especially deep learning methods used in image processing, has promoted connectomic data analysis to a new era. Even though, the performance of the state-of-the-art methods is still fall behind the demand of researchers. To alleviate this situation, here we first introduce a new annotated Ultra-high Resolution Image Segmentation dataset for the Cell membrane, called U-RISC, which is the largest annotated Electron Microscopy (EM) dataset for the cell membrane with multiple iterative annotations and the resolution of 2.18nm/pixel. Then we reveal the performance of existing deep learning segmentation methods on U-RISC through an open competition. The performance of participants appears to have a huge gap with human-level, however, the results of same methods on ISBI 2012, a smaller EM dataset, are near-human. To further explore the differences between the performance of two datasets, we analyze the neural networks with attribution analysis and uncover the larger decision-making area in the segmentation of U-RISC. Our work provides a new benchmark data for EM cell membrane segmentation and proposes some perspectives in deep learning segmentation algorithms.

No abstract

Motivation Cell membrane segmentation in electron microscopy (EM) images is a crucial step in EM image processing. However, while popular approaches have achieved performance comparable to that of humans on low-resolution EM datasets, they have shown limited success when applied to high-resolution EM datasets. The human visual system, on the other hand, displays consistently excellent performance on both low and high resolutions. To better understand this limitation, we conducted eye movement and perceptual consistency experiments. Our data showed that human observers are more sensitive to the structure of the membrane while tolerating misalignment, contrary to commonly used evaluation criteria. Additionally, our results indicated that the human visual system processes images in both global-local and coarse-to-fine manners. Results Based on these observations, we propose a computational framework for membrane segmentation that incorporates these characteristics of human perception. This framework includes a novel evaluation metric, the perceptual Hausdorff distance (PHD), and an end-to-end network called the PHD-guided segmentation network (PS-Net) that is trained using adaptively tuned PHD loss functions and a multiscale architecture. Our subjective experiments showed that the PHD metric is more consistent with human perception than other criteria, and our proposed PS-Net outperformed state-of-the-art methods on both low and high resolution EM image datasets as well as other natural image datasets. Availability The code and dataset can be found at https://github.com/EmmaSRH/PS-Net. Supplementary information Supplementary Information for this article is available online.