Towards Unifying Anatomy Segmentation: Automated Generation of a Full-Body CT Dataset

Multi-energy reconstructions have become an important research field in computed tomography in recent years. Since modern reconstruction and postprocessing techniques often employ deep learning strategies, there is a high need for large, diverse and adaptable multi-energy datasets. Therefore, this work proposes a straightforward pipeline for the generation of multi-energy cone-beam CT projection data based on the established XCAT software phantom with arbitrary desired X-ray spectra. We evaluate the effort and time required for dataset generation and utilize the generated data for model-based iterative reconstruction exemplarily. This approach provides an understanding of the current pipeline’s bottlenecks while demonstrating its suitability in producing high-quality projection datasets and reconstructions. Thus, we contribute to open knowledge on generation of large multi-energetic CT datasets for deep learning purposes.

Cardiac C-arm CT imaging delivers a tomographic region-of-interest reconstruction of the patient's heart during image guided catheter interventions. Due to the limited size of the flat detector a volume image is reconstructed, which is truncated in the cone-beam (along the patient axis) and the fan-beam (in the transaxial plane) direction. To practically address this local tomography problem correction methods, like projection extension, are available for first pass image reconstruction. For second pass correction methods, like metal artefact reduction, alternative correction schemes are required when the field of view is limited to a region-of-interest of the patient. In classical CT imaging metal artefacts are corrected by metal identification in a first volume reconstruction and generation of a corrected projection data set followed by a second reconstruction. This approach fails when the metal structures are located outside the reconstruction field of view. When a C-arm CT is performed during a cardiac intervention pacing leads and other cables are frequently positioned on the patients skin, which results in propagating streak artefacts in the reconstruction volume. A first pass approach to reduce this type of artefact is introduced and evaluated here. It makes use of the fact that the projected position of objects outside the reconstruction volume changes with the projection perspective. It is shown that projection based identification, tracking and removal of high contrast structures like cables, only detected in a subset of the projections, delivers a more consistent reconstruction volume with reduced artefact level. The method is quantitatively evaluated based on 50 simulations using cardiac CT data sets with variable cable positioning. These data sets are forward projected using a C-arm CT system geometry and generate artefacts comparable to those observed in clinical cardiac C-arm CT acquisitions. A C-arm CT simulation of every cardiac CT data set without cables served as a ground truth. The 3D root mean square deviation between the simulated data set with and without cables could be reduced for 96% of the simulated cases by an average of 37% (min −9%, max 73%) when using the first pass correction method. In addition, image quality improvement is demonstrated for clinical whole heart C-arm CT data sets when the cable removal algorithm was applied.

Purpose: Radiotherapy treatment planning studies often require the use of large patient CT datasets to extract conclusions of statistical significance. However, due to various reasons such as difficulties in acquiring a larger set of CT scans and in segmenting organs in all the images, real studies are usually performed using very limited datasets. The aim of this study is to develop a novel method for generating large datasets of realistic patient geometries for treatment planning studies using a computerized phantom. Method and Materials: A NURBS-based cardiac-torso (NCAT) phantom was built based upon data from the female Chinese Visible Human (CVH) datasets. The NURBS control points on the organs surface were deformed to match the organ surface obtained from the limited daily cone beam CT (CBCT) dataset of each patient undergoing adaptive radiotherapy. A principal component analysis (PCA) of control point deformations was performed for the individual patient and each geometry was fit to a combination of the principal components with their corresponding weighting factors. A statistical analysis of the weighting factors was performed, and a new larger set of statistically equivalent weighting factors can be constructed, which will result in a larger geometry dataset for the patient. Results: The NCAT pelvis phantom was developed based on the segmentation of organs in the Chinese Visible Human pelvic dataset. The method for generating new datasets was applied to 20 patients undergoing adaptive radiotherapy and a variety of realistic deformed pelvic geometries was developed. Conclusion: We present a novel method for automatic generation of large datasets of patient geometries from a set of limited image datasets. The new geometries should have the same statistical uncertainties as the original datasets, however much smaller random uncertainties, and can be used in the future for adaptive radiotherapy planning studies.

Deformable registration has been one of the pillars of biomedical image computing. Conventional approaches refer to the definition of a similarity criterion that, once endowed with a deformation model and a smoothness constraint, determines the optimal transformation to align two given images. The definition of this metric function is among the most critical aspects of the registration process. We argue that incorporating semantic information (in the form of anatomical segmentation maps) into the registration process will further improve the accuracy of the results. In this paper, we propose a novel weakly supervised approach to learn domain-specific aggregations of conventional metrics using anatomical segmentations. This combination is learned using latent structured support vector machines. The learned matching criterion is integrated within a metric-free optimization framework based on graphical models, resulting in a multi-metric algorithm endowed with a spatially varying similarity metric function conditioned on the anatomical structures. We provide extensive evaluation on three different datasets of CT and MRI images, showing that learned multi-metric registration outperforms single-metric approaches based on conventional similarity measures.

Voxel-based morphometry (VBM) is a neuroimaging technique that investigates focal differences in brain anatomy. The core process of VBM is segmenting the brain into grey matter, white matter, and cerebrospinal fluid, warping the segmented images to a template space and smoothing. Thereafter, statistical analysis is performed on the basis of the general linear model. Although the basis of VBM is constant, the algorithm has been changed. Classical VBM simply employed anatomical normalization, segmentation, and smoothing. This changed to optimized VBM, which normalized the brain using parameters derived from grey matter image normalization, cleaned up non-brain tissue images, and utilized Jacobian modulation. Further, unified segmentation-a probabilistic framework that enables image registration, tissue classification, and bias correction to be combined within the same generative model-was introduced. The DARTEL algorithm then improved the accuracy of image registration. Currently, researchers can use an extension of unified segmentation with some features such as an improved registration model, extended set of tissue probability maps, or more robust initial affine registration. Those who utilize VBM must pay attention to the choice of VBM algorithm, as data interpretation differs with each algorithm.

Inverse planning based on intensity‐modulated radiation therapy (IMRT) requires the delineation of target volumes and normal critical organs. Anatomy segmentation becomes an important step in modern radiation therapy. Unfortunately, anatomy segmentation is traditionally done by a manual contouring process, which can be labor‐intensive and subject to inter‐observer variations. When many anatomical structures are to be contoured in 3D or 4D CT datasets, the manual contouring process can be a bottleneck for IMRT planning, especially in adaptive radiotherapy which involves multiple, sequential re‐planning. Recently, deformable image registration demonstrates its advantage for auto‐segmentation. The basic assumption is that patientˈs anatomy can be deformably mapped from a previously defined anatomy configuration — the atlas. This reference anatomy can be the same patient (in adaptive radiotherapy) or a different patient within the same class (or a patient with the same type of cancer). Deformable image registration provided a voxel‐by‐voxel transformation field, which can be used to map contours (or the labeled volumes) from the reference patient to the new image. Atlas‐based auto‐segmentation has a special advantage in image‐guided adaptive radiotherapy, because the original atlas contains both target volume and normal structures, and is already defined in the original treatment plan, making treatment adaptation a much simple process. In this presentation, we will discuss our clinical experience for using atlas‐based segmentation for intra‐object (the same patient) contour propagation and inter‐object (a different patient) auto‐ segmentation. Validation of auto‐segmentation is also an important step to be discussed.Educational Objectives:1. Describe deformable image registration for auto‐segmentation2. Understand the achievable accuracy and validation methods for anatomy‐segmentation.3. Illustrate applications of atlas‐based auto‐segmentation in radiation therapy

To develop a software tool that analyzes the anatomy of the portal vein branches and assigns segmental and subsegmental branches according to Couinaud's classification system and to evaluate its accuracy. The algorithm was developed in C++ on a PC. The algorithm recognizes the three major branching patterns of the portal vein. Segmental and subsegmental branches are assigned to 8 segments following Couinaud and encoded by 8 colors. The software was evaluated using CT data sets of 39 patients. After the individual segmental anatomy of each patient was determined by an experienced radiologist, automatic classification was performed and the results were compared on a branch by branch basis. The numbering was accurate according to Couinaud's system in 358 of 409 segmental and subsegmental branches (88 %). The assignment failed in 51 of 409 branches due to unexpected anatomy or software problems. Automatic classification of portal vein branches and their appendant parenchyma is possible. The automatic designation of liver segments enables the three-dimensional visualization of the segmental anatomy. In the future, automatic analysis might facilitate the reporting and communication of CT findings.

Deep convolutional neural networks (CNNs) have shown impressive performance in anatomy segmentation that are close to the state of the art atlas-based segmentation method. On one hand CNNs have 20x faster predictions than atlas-based segmentation. However, one of the main holdbacks of CNN’s advancement is that it’s training requires large amount of annotated data. This is a costly hurdle as annotation is time consuming and requires expensive medical expertise. The goal of this work is to reach state of the art segmentation performance using the minimum amount of expensive manual annotations. Recent studies show that auxiliary segmentations can be used together with manual annotations to improve CNN learning. To make this learning scheme more effective, we propose an image selection algorithm that wisely chooses images for manual annotation for producing more accurate auxiliary segmentations and a quality control algorithm that excludes poor quality auxiliary segmentations from CNN training. We perform extensive experiments over chest CT dataset by varying the number of manual annotations used for atlas-based methods and by varying the number of auxiliary segmentations to train the CNN. Our results show that CNN trained with auxiliary segmentations achieve higher dice of 0.76 vs 0.58 when trained with few accurate manual segmentations. Moreover, training with 100 or more auxiliary segmentations, the CNN always outperforms atlas-based method. Finally, when carefully selecting single atlas for producing auxiliary segmentations and controlling the quality of auxiliary segmentations, the trained CNN archives high average dice of 0.72 vs 0.62 when using a randomly selected image for manual annotation with all auxiliary segmentations.

At present, deep learning has been widely adopted in medical image processing. However, the current deep neural networks depend on a large number of labeled training data, but medical images segmentation tasks often suffer from the problem of small quantity of labeled data because labeling medical images is a very expensive and time-consuming task. In order to overcome this difficulty, this paper proposes a new image augmentation strategy based on statistical shape model and three-dimensional thin plate spline, which can generate many simulated images from a small number of real images. Firstly, the shape information of the real labeled images is modeled with the statistical shape model, and a series of simulated shapes are generated by sampling from this model. Secondly, the simulated shapes are filled with texture using three-dimensional thin plate spline to generate the simulated images. Finally, the simulated images and the real images are used together for training deep neural networks. The proposed framework is a general data augmentation method that can be used in any anatomical structure segmentation tasks with any deep neural network architecture. We used two different datasets, including prostate MRI dataset and liver CT dataset, and used two different deep network structures, including multi-scale 3D Convolutional Neural Networks (multi-scale 3D CNN) and U-net. The experimental results showed that the proposed data augmentation strategy can improve the accuracy of existing segmentation algorithms based on deep neural networks.

To explore the feasibility of a fully automated workflow for whole-body volumetric analyses based on deep reinforcement learning (DRL) and to investigate the influence of contrast-phase (CP) and slice thickness (ST) on the calculated organ volume. This retrospective study included 431 multiphasic CT datasets—including three CP and two ST reconstructions for abdominal organs—totaling 10,508 organ volumes (10,344 abdominal organ volumes: liver, spleen, and kidneys, 164 lung volumes). Whole-body organ volumes were determined using multi-scale DRL for 3D anatomical landmark detection and 3D organ segmentation. Total processing time for all volumes and mean calculation time per case were recorded. Repeated measures analyses of variance (ANOVA) were conducted to test for robustness considering CP and ST. The algorithm calculated organ volumes for the liver, spleen, and right and left kidney (mean volumes in milliliter (interquartile range), portal venous CP, 5 mm ST: 1868.6 (1426.9, 2157.8), 350.19 (45.46, 395.26), 186.30 (147.05, 214.99) and 181.91 (143.22, 210.35), respectively), and for the right and left lung (2363.1 (1746.3, 2851.3) and 1950.9 (1335.2, 2414.2)). We found no statistically significant effects of the variable contrast phase or the variable slice thickness on the organ volumes. Mean computational time per case was 10 seconds. The evaluated approach, using state-of-the art DRL, enables a fast processing of substantial amounts irrespective of CP and ST, allowing building up organ-specific volumetric databases. The thus derived volumes may serve as reference for quantitative imaging follow-up.

Purpose: Traditional extended SSD total body irradiation (TBI) techniques can be problematic in terms of patient comfort and/or dose uniformity. This work aims to develop a comfortable TBI technique that achieves a uniform dose distribution to the total body while reducing the dose to organs at risk for complications. Methods: To maximize patient comfort, a lazy Susan-like couch top immobilization system which rotates about a pivot point was developed. During CT simulation, a patient is immobilized by a Vac-Lok bag within the body frame. The patient is scanned head-first and then feet-first following 180° rotation of the frame. The two scans are imported into the Pinnacle treatment planning system and concatenated to give a full-body CT dataset. Treatment planning matches multiple isocenter volumetric modulated arc (VMAT) fields of the upper body and multiple isocenter parallel-opposed fields of the lower body. VMAT fields of the torso are optimized to satisfy lung dose constraints while achieving a therapeutic dose to the torso. The multiple isocenter VMAT fields are delivered with an indexed couch, followed by body frame rotation about the pivot point to treat the lower body isocenters. The treatment workflow was simulated with a Rando phantom, and the plan was mapped to a solid water slab phantom for point- and film-dose measurements at multiple locations. Results: The treatment plan of 12Gy over 8 fractions achieved 80.2% coverage of the total body volume within ±10% of the prescription dose. The mean lung dose was 8.1 Gy. All ion chamber measurements were within ±1.7% compared to the calculated point doses. All relative film dosimetry showed at least a 98.0% gamma passing rate using a 3mm/3% passing criteria. Conclusion: The proposed patient comfort-oriented TBI technique provides for a uniform dose distribution within the total body while reducing the dose to the lungs.

The aim of this study was to evaluate the feasibility of implementing volumetric modulated arc therapy (VMAT)-based techniques and extended CBCT image guidance for total body irradiation (TBI) treatment using a conventional linear accelerator. Patients eligible for TBI between November 2016 and December 2024 were included in the analysis. Patients received a total dose of 4-12Gy, given in six fractions within 3days, two fractions/day with 6h minimum interval between fractions, or 2Gy in one fraction, depending on the clinical indication. During the initial phase of the protocol, PET-CT imaging was used to obtain full-body CT datasets. Subsequently, CT simulation was performed using a multislice Siemens CT scanner available in the Radiation Oncology Department. In this setting, two CT studies were acquired per patient, one extending from the pelvis to the head (upper scan) and a second from the pelvis to the feet (lower scan), which were merged into a single dataset for treatment planning. Dosimetric planning was performed using a multi-isocenter approach with the Eclipse™ treatment planning system, employing volumetric modulated arc therapy (VMAT) to achieve the prescribed dose distribution. In the initial stage of the treatment program, treatments were delivered on a Varian CLINAC DHX linear accelerator. Following its decommissioning, treatment delivery was transitioned to Varian TrueBeam linear accelerators (models SN3790 and SN2137). Treatment delivery, including verification and patient positioning, was performed sequentially, beginning with the upper body followed by the lower body. Image guidance was initially based on kV-MV imaging and was later replaced by extended-field cone-beam computed tomography (CBCT), which was registered to the simulation CT to enable automated setup corrections. Dosimetric parameters and setup verification metrics were subsequently analyzed. Between November 2016 to December 2024, 27 patients fulfilled the inclusion criteria. All scheduled sessions were completed, amounting to a total of 148 treatment fractions. The average number of isocenters used to generate the treatment plans was 7,11 (6-12). The mean lung dose was 10.12Gy (range 8.97-11.07Gy). Dose homogeneity achieved across all sessions was 1.24 (1.11-1.41). After image acquisition, mean setup corrections were 0.06cm lateral (range 0.00-2.00cm), 0.26cm vertical (0.00-2.00cm), and 0.03cm longitudinal (0.00-1.30cm) in head-first plan. Laterally, 0.26cm (range: 0.00-2.00cm) vertically, and 0.03cm (range: 0.00-1.30cm) longitudinally in feet first plan. The average duration of each session, from the first image acquisition to the completion of the final field, was 87min (range 60-159). Our study demonstrates that VMAT-based TBI is a feasible and promising alternative to conventional 2D-TBI, providing improved dose homogeneity, enhanced organ sparing and with reproducibility comparable to previously reported HT systems. These findings support the integration of VMAT techniques on conventional LINACs for TBI treatments, although further prospective studies are needed to confirm long-term clinical benefits.

In this paper, we propose a novel 3D automatic anatomy segmentation method based on the synergistic combination of active appearance models (AAM), live wire (LW) and graph cut (GC). The proposed method consists of three main parts: model building, initialization and segmentation. For the model building part, an AAM model is constructed and the LW cost function is trained. For the initialization part, an improved iterative model refinement algorithm is proposed for the AAM optimization, which synergistically combines the AAM and LW method (OAAM). And a multi-object strategy is applied to help the object initialization. A pseudo 3D initialization strategy is employed to segment the organs slice by slice via multi-object OAAM method. The model constraints are applied to the initialization result. For the segmentation part, the object shape information generated from the initialization step is integrated into the GC cost computation. And an iterative GCOAAM method is proposed for object delineation. This method is a general method and can be applied to any organ segmentation. The proposed method was tested on the clinical liver and kidney CT data sets. The results showed the following: (a) an overall segmentation accuracy of true positive fraction>93.5%, and false positive fraction<0.2% can be achieved. (b) The initialization performance is improved by combining the AAM and LW. (c) The multi-object strategy greatly helps the initialization due to inter-object constraints.