Exploring Flood Filling Networks for Instance Segmentation of XXL-Volumetric and Bulk Material CT Data

Abstract

XXL-Computed Tomography (XXL-CT) is able to produce large scale volume datasets of scanned objects such as crash tested cars, sea and aircraft containers or cultural heritage objects. The acquired image data consists of volumes of up to and above $$\hbox {10,000}^{3}$$ 10,000 3 voxels which can relate up to many terabytes in file size and can contain multiple 10,000 of different entities of depicted objects. In order to extract specific information about these entities from the scanned objects in such vast datasets, segmentation or delineation of these parts is necessary. Due to unknown and varying properties (shapes, densities, materials, compositions) of these objects, as well as interfering acquisition artefacts, classical (automatic) segmentation is usually not feasible. Contrarily, a complete manual delineation is error-prone and time-consuming, and can only be performed by trained and experienced personnel. Hence, an interactive and partial segmentation of so-called “chunks” into tightly coupled assemblies or sub-assemblies may help the assessment, exploration and understanding of such large scale volume data. In order to assist users with such an (possibly interactive) instance segmentation for the data exploration process, we propose to utilize delineation algorithms with an approach derived from flood filling networks. We present primary results of a flood filling network implementation adapted to non-destructive testing applications based on large scale CT from various test objects, as well as real data of an airplane and describe the adaptions to this domain. Furthermore, we address and discuss segmentation challenges due to acquisition artefacts such as scattered radiation or beam hardening resulting in reduced data quality, which can severely impair the interactive segmentation results.

3D segmentation of plant root systems using spatial pyramid pooling and locally adaptive field-of-view inference

“…Examples of the first category include the 'RootForce' algorithm (Gerth et al, 2021), which solves this task by analyzing the curvature of the gray value profile of the root voxels using classical image processing and analysis approaches. Using the specific mean gray values of the two object types 'root' and 'nonroot', RootForce sorts out all objects which are not in a specific gray value range and thereby generates a mask to indicate possible root voxels.…”

“…This is important since a classical ground truth via root excavation and washing, results in a loss of fine root structure and interconnectivity of the whole root system. To this end, the 3D segmentation of the root structures was mainly performed by analytical algorithms based on classical image processing and image analysis methods [e.g., Mairhofer et al, 2011;Flavel et al, 2012;Mairhofer et al, 2015;Flavel et al, 2017;Gao et al, 2019;Soltaninejad et al, 2020;Gerth et al, 2021;Phalempin et al, 2021;Ferreira et al, 2022;Lucas & Vetterlein, 2022), which, however, are not able to detect roots on all scales equally. See Figure 1 for an example which contains challenging small and fine roots.…”

“…To overcome the limitations of the current analytical methods, we propose an improved deep neural network based approach, which is (a) able to adapt to different feature scales of the roots to be delineated, (b) can be trained with noisy information, and (c) furthermore is able to efficiently handle large scale 3D data sets of roots during inference. We make use of a classical deep convolutional neural network (DCNN) architecture which is enhanced by a 'Spatial Pyramid Pooling' (SPP) layer (He et al, 2014) (see Section 4) to gain scale invariance for root segmentation by supporting arbitrarily sized 'field of views' (FoVs) (see Section 3.4) as input sample and is furthermore using a weakly supervised training scheme (Khoreva et al, 2017) (see Section 5).…”

Bin-scanning: Segmentation of X-ray CT volume of binned parts using Morse skeleton graph of distance transform

“…For industrial X-ray CT, a recent study [47] applied the flood filling network [48] to segment a large CT volume with 10,000 3 voxels. However, this method aims to segment the regions of parts made of different materials and cannot be used for the parts made of the same material.…”

Thickness-Driven Sheet Metal Segmentation of CT-Scanned Body-in-White

“…Therefore, it is challenging to segment the CT volume using conventional manual or semiautomatic segmentation systems. Researchers have thoroughly investigated this problem, suggesting solutions such as automation using machine learning and scalability for enormous quantities of data [3,4].…”