Universal Soldier: A deep neural net for unsupervised 3D segmentation of tomographic images of bones

通用士兵深度网络是一个用于图像分析的人工卷积神经网络。“通用战士”（USDNet）将是Natalie Reznikov教授与目标研究系统公司（ORS）合作项目的成果，后者的产品是蜻蜓软件（Dragonfly），用于全面的3D图像分析。目的是设计和训练一个深度人工神经网络（通用士兵），该网络将能够对通过各种基于x射线的方法获得的骨骼3D图像进行无监督分割。目前，图像分割——即在3D中识别和准确标记相关特征——是生物成像的一个瓶颈，这主要是因为生物物体（如骨骼）的层次复杂性很高，加上累积的3D数据占用空间很大。3D数据集的自动、无偏分割将消除这些限制，提高定量图像分析的精度，具有高通量。从2020年到2022年，我们收集了大量使用x射线计算机断层扫描（CT）扫描仪获得的各种动物（包括人类）骨骼的3D断层图像库，分辨率范围从1 μ m到60 μ m，并具有各种自然产生的人工制品。该库目前的结构是一个SQL存储库，包含大约2tb的原始图像，以及经过专业处理的原始数据子样本（训练数据，或“基本事实”，约5%）。有了这个库，作为这个拟议项目的一部分，我们现在将设计和训练USDNet，它将能够在任何扫描中识别骨骼元素，并产生无监督的、高保真的自动分割。这个万能战士将成为图像分析软件蜻蜓的一部分，骨骼生物学家和生物成像研究人员可以免费使用。这将在生命科学中普及人工智能辅助方法，将使定量3D图像分析快速、简化且不受认知偏差的影响。就像自动驾驶汽车今天已经成为现实一样，我们相信，使用预先训练过的通用士兵深度网络的自动分割将在未来改变生物成像。

The Universal Soldier Deep Net is an artificial convolutional neural net for image analysis. The Universal Soldier, or USDNet, will be the output of this collaborative project between Prof. Natalie Reznikov and Object Research Systems (ORS) Inc. (Montréal), whose product is the software Dragonfly for comprehensive 3D image analysis. The purpose is to design and train a deep artificial neural network (the Universal Soldier) that will be capable of unsupervised segmentation of 3D images of bones as acquired by various X-ray-based methods. Currently, image segmentation - i.e. the identification and accurate tagging of relevant features in 3D - is a bottleneck of bioimaging, largely because of the high degree of hierarchical complexity in biological objects (such as bones), together with the large footprint of accrued 3D data. Automated, unbiased segmentation of 3D datasets would abolish these limitations and increase the precision of quantitative image analysis, with high throughput. From 2020-22, we collected a vast library of 3D tomographic images of bones of various animals (including humans), acquired using X-ray computed tomography (CT) scanners, with resolutions ranging from 1 µm to 60 µm, and with a variety of naturally occurring artifacts. The library is currently structured as an SQL repository and contains about 2 TB of raw images, as well as expertly processed subsamples of raw data (training data, or "ground truth", about 5%). Having this library, as part of this proposed project we will now design and train the USDNet that will be able to recognize skeletal elements in any scan and produce unsupervised, high-fidelity automated segmentation. This Universal Soldier will become part of the image analysis software Dragonfly available to skeletal biologists and bioimaging researchers free of charge. This will popularize artificial intelligence-aided methodologies in the life sciences, will make quantitative 3D image analysis fast, streamlined and immune to cognitive biases. Like self-driving cars have become a reality today, automated segmentation using a pre-trained Universal Soldier Deep Net we believe will transform bioimaging tomorrow.