PaperlayerView exact revision on arXiv

Source-linked AI summary

Deep learning-based synthetic-CT generation in radiotherapy and PET: a review

Maria Francesca Spadea, Matteo Maspero, Paolo Zaffino, Joao Seco

arXiv:2102.02734v2physics.med-phcs.LGeess.IV

TL;DR

Deep learning-based synthetic-CT generation has been studied across MR treatment planning, CBCT-guided adaptive radiotherapy, and PET attenuation correction, but its clinical readiness requires systematic assessment. This review categorizes and compares published methods, finding feasibility in radiotherapy and performance comparable to standard MR attenuation-correction methods for PET.

  • Problem

    Daily CBCT has not commonly been used for online plan adaptation, motivating assessment of synthetic-CT methods across clinical applications.

    Source: §I · p. 3

  • Method

    The review groups synthetic-CT generation studies by clinical application and compiles their techniques, key findings, and pertinent characteristics.

    Source: §II · pp. 5–6

  • Results

    Dose differences from CT-based planning were below 1% in radiotherapy contexts, while synthetic-CT showed better performance for PET attenuation correction than standard MRAC methods.

    Source: §V · p. 32

  • Takeaways & Limitations

    Synthetic-CT generation seems feasible for radiotherapy at head-and-neck, brain, and pelvic sites, with promising performance for PET attenuation correction.

    Source: §V · p. 32

    Limited field of view may hinder plan calculation for replanning purposes.

    Source: §IV · p. 26

Abstract

from arXiv · show

Recently, deep learning (DL)-based methods for the generation of synthetic computed tomography (sCT) have received significant research attention as an alternative to classical ones. We present here a systematic review of these methods by grouping them into three categories, according to their clinical applications: I) To replace CT in magnetic resonance (MR)-based treatment planning. II) Facilitate cone-beam computed tomography (CBCT)-based image-guided adaptive radiotherapy. III) Derive attenuation maps for the correction of positron emission tomography (PET). Appropriate database searching was performed on journal articles published between January 2014 and December 2020. The DL methods' key characteristics were extracted from each eligible study, and a comprehensive comparison among network architectures and metrics was reported. A detailed review of each category was given, highlighting essential contributions, identifying specific challenges, and summarising the achievements. Lastly, the statistics of all the cited works from various aspects were analysed, revealing the popularity and future trends, and the potential of DL-based sCT generation. The current status of DL-based sCT generation was evaluated, assessing the clinical readiness of the presented methods.

I. Introduction · II. Material and Methods

The review motivates deep learning-based synthetic CT generation by outlining unmet needs across MR-only radiotherapy, CBCT-guided adaptation, and PET attenuation correction. It systematically evaluates eligible methods, architectures, metrics, trends, challenges, and clinical readiness using a PRISMA-guided literature search.

I. Introduction

  • CT remains central to radiotherapy because it provides accurate patient geometry and electron-density conversion for dose calculation [7].MRI adds superior soft-tissue contrast for tumour and organ-at-risk delineation [9,10], while PET supplies functional and metabolic information [8].Source: §I · p. 2
  • MR-only radiotherapy could reduce registration errors, workflow burden, ionising-radiation exposure, costs, and workload, but requires accurate tissue attenuation information for dose calculation [12,21].Synthetic CT methods convert MRI into CT-equivalent representations to address this obstacle.Source: §I · p. 2

I. INTRODUCTION

  • CBCT artefacts from scatter noise and truncated projections have limited daily online plan adaptation, whereas CBCT-to-CT conversion could enable accurate dose computation and improve IGART [29–31].The cited artefacts include shading, streaking, and cupping.Source: §I · p. 3
  • Synthetic CT estimation also supports PET attenuation correction, addressing the difficulty of deriving reliable attenuation maps in PET/MRI and from uncorrected standalone PET.Accurate PET quantification requires a reliable photon attenuation-correction map, usually derived from CT.Source: §I · p. 3
  • The review addresses a detailed gap in prior literature by systematically summarising DL-based sCT developments across applications, imaging protocols, architectures, achieved performance, trends, challenges, and clinical readiness.Earlier reviews covered narrower topics, whereas this survey explicitly organises contributions by application and discusses clinical readiness.Source: §I · p. 3

II. Material and Methods

  • The systematic review followed PRISMA guidelines and searched PubMed, Scopus, and Web of Science for studies published from January 2014 through December 2020.Duplicate removal, eligibility screening, and citation searching followed the database search.Source: §II · p. 4
  • Eligibility covered external-beam photon or proton radiotherapy and PET attenuation correction using sCT generated from MRI, CBCT, or PET, while excluding brachytherapy, conventional machine learning, dual-energy CT, direct PET attenuation-map estimation, and conference proceedings.Each included study was assigned to a clinical application category.Source: §II · p. 4

I MR-only RT; … II.A. Deep learning for image synthesis

The review organizes DL-based sCT generation across MR-only radiotherapy, CBCT-based adaptive radiotherapy, and PET attenuation correction, then explains architectures, configurations, evaluation, and synthesis mechanisms. It emphasizes CNN-based image-to-image translation, including U-nets, GANs, and cycle-GANs for unpaired training.

I MR-only RT;

II. MATERIAL AND METHODS

  • The methods framework includes paired or unpaired data, independent test-set inference or ensembles, publication-year statistics, anatomical-region distributions, and clinical-feasibility discussion.Source: III PET attenuation correction · p. 4, §II · pp. 5–6
  • The review distinguishes 2D+ networks combined across views, multi-2D networks receiving slices from different views, 2.5D networks using neighbouring slices, and 3D volume-based inputs.Source: §II · p. 5

III PET attenuation correction.

  • For each application, tables summarize published techniques, key findings, anatomical sites, patient numbers, imaging protocols, architectures, configurations, data pairing, treatments, and evaluation metrics.Source: §II · pp. 5–6

II.A. Deep learning for image synthesis

Deep learning for image synthesis

  • Cycle-GANs enable unpaired image-to-image translation through forward and backward mappings with cycle-consistency losses minimizing differences between original and reconstructed images.Source: Deep learning for image synthesis · p. 7

II.B. Metrics

The review groups sCT evaluation metrics into image similarity, geometric accuracy, and task-specific measures. Their interpretation depends on image alignment, application-specific choices, and reporting parameters.

II.B. Metrics

  • Metrics are organized into image similarity, geometric accuracy, and task-specific categories, as summarized in Table 1.Image-similarity measures include M(A)E or MAPE, PSNR, SSIM, CC, NCC, MSE, and RMSE.Source: II.B. Metrics · p. 8, Table 1 · p. 8

Metrics

  • Voxel-wise image-similarity metrics compare aligned sCT and target CT data after translation, rigid, affine, or deformable registration.MAE and ME characterize random and systematic error, MSE and RMSE emphasize larger errors, PSNR evaluates synthesis noise, and SSIM reflects structural similarity using human-visual-system characteristics59.Source: II.B. Metrics · p. 8, Metrics · p. 9
  • Geometric accuracy compares segmented structures between CT and sCT using segmentation metrics such as DSC, HD, and mean absolute surface distance.Structures may be manually or automatically segmented, with thresholding and morphological operations used to create binary masks; metric selection should be assessed on an application basis62.Source: Metrics · p. 9
  • Task-specific radiotherapy metrics include dose difference, dose pass rate, gamma pass rate, dose-volume-histogram differences, and proton range-shift error.Dose comparisons may use absolute or relative differences, while gamma results depend on dose and distance criteria, analysis mode, dose grid size, and voxel resolution64,65.Source: Metrics · pp. 9–10
  • For PET attenuation correction, evaluation usually reports signed and unsigned relative PET reconstruction error.In proton radiotherapy, range-shift error can be reported as either an absolute difference or a relative percentage along the beam direction68.Source: Metrics · p. 10

II. MATERIAL AND METHODS II.B. Metrics

Metric comparisons can be challenging because identical metrics may be calculated using different regions of interest and implementation procedures. The review therefore highlights these differences and considers their possible impact.

II.B. Metrics

  • Different regions of interest can make comparisons difficult even when studies calculate the same metric.Examples include whole predicted volumes, volumes of interest, and cropped volumes.Source: §II · p. 11
  • Metric implementations can change across studies, so the review highlights these differences and speculates on their impact.This variability affects the interpretation of performance comparisons.Source: §II · p. 11
  • Gamma analysis varies with dose-difference and distance-to-agreement criteria, including γ3%,3mm (γ3), γ2%,2mm (γ2), and γ1%,1mm (γ1).Implementation may also differ in the dose-threshold ROI and between 2D and 3D algorithms.Source: §II · p. 11

III. Results

The review identified 83 eligible papers, with most studies addressing MR-only radiotherapy and publication numbers generally increasing over time. Across applications, investigations predominantly involved adult cohorts, paired regression-based 2D networks, and GAN or U-net architectures, with brain, pelvis, and head-and-neck regions most frequently studied.

III. Results

  • 83 eligible papers remained after duplicate removal and content screening from searches yielding 91 PubMed, 98 Scopus, and 218 Web of Science records.The first included work was published in 2017 after conference papers were excluded, and publication numbers generally increased over time.Source: §III · p. 11
  • MR-only radiotherapy comprised 51 studies (61.4%), compared with 15 (18.1%) for CBCT-to-CT adaptive radiotherapy and 17 (20.5%) for PET attenuation correction.CBCT-to-CT and PET attenuation-correction publications were stable in the last years, unlike the broader upward trend.Source: §III · p. 11
  • Brain, pelvis, and head-and-neck were the most investigated anatomical regions, covering approximately 80% of MR-only studies and over 75% of CBCT-to-CT and PET studies.Head-and-neck and pelvic sites were most explored for CBCT-to-CT, while head-and-neck followed by pelvic sites dominated PET attenuation-correction studies.Source: §III · p. 11
  • Most studies included fewer than 50 patients across all three categories, and adult cohorts predominated.Paediatric datasets were described as more heterogeneous for network training; feasibility was first investigated for PET attenuation correction in 79 patients and later for photon and proton applications.Source: §III · pp. 11, 13

III. RESULTS

  • Nearly all models performed input-to-sCT regression, while only two studies used segmentation into predefined classes; paired training predominated, with unpaired training in 13/83 articles.Four studies directly compared paired and unpaired training71,79,80,81.Source: §III · p. 13
  • Across all applications, 2D networks were used about 61% of the time, followed by 3D (24%), 2.5D (10%), and 2D+ (6%), while GANs and U-nets were the most common architectures.GANs appeared 45 times and U-nets 36 times; all investigations used registration for sCT evaluation except Xu et al.81 and Fetty et al.84, which used unpaired metrics such as FID.Source: §III · p. 13

III.A. MR-only radiotherapy

In MR-only radiotherapy, DL-based sCT methods predominantly used 2D GAN or U-net architectures and generally achieved strong image- and dose-based performance. Results depended on MRI inputs, dimensionality, training pairing, and evaluation practices, with some contradictory architecture comparisons and limited robustness assessment.

MR-only radiotherapy

  • Most MRIs were acquired at 1.5 T or 3 T, typically with site-specific T1 GRE or T2 TSE sequences, but more than ten studies omitted sequence or field-strength details.Most studies used a single MRI sequence, while preprocessing commonly included intensity normalization, histogram matching, linear rescaling, bias-field correction, or intensity homogenization.Source: III.A. MR-only radiotherapy · p. 14
  • Multiple MRI contrasts generally improved sCT generation, with Gadolinium-enhanced T1 GRE yielding the lowest MAE and highest PSNR in brain imaging and multiple Dixon images benefiting pelvic performance.Combining T1 and T2 sequences also outperformed single-sequence models.Source: III.A. MR-only radiotherapy · p. 14
  • 2D+ inference was beneficial, while reported comparisons between 3D and 2D U-nets were conflicting.Fu et al.82 found that a modified 3D U-net outperformed a 2D U-net, whereas Neppl et al.83 reported underperformance on image similarity metrics and photon- and proton-dose differences.Source: III.A. MR-only radiotherapy · pp. 14, 18
  • Paired training generally produced better image similarity and lower dose differences than unpaired training, while mixed paired/unpaired training and structure-constrained losses improved unpaired performance.Good registration was important for paired GANs to outperform unpaired training.Source: III.A. MR-only radiotherapy · p. 18
  • Among employed architectures, GANs accounted for approximately 55% of studies and U-nets approximately 35%, while ensembles outperformed single models and were more robust across different scanners and centres.U-nets and GANs could achieve similar image- and dose-based performances.Source: III.A. MR-only radiotherapy · p. 18
  • Average dose differences were <1% and γ2%,2mm pass rates exceeded 95%, except in one study, although metric calculation was inconsistent across anatomical sites.Thirty studies evaluated sCT-based dose calculation, including photon, proton, or both modalities; only four assessed robustness across multiple conditions.Source: III.A. MR-only radiotherapy · pp. 18–19

III.B. CBCT-to-CT generation

CBCT-to-CT synthesis is the newest CT-synthesis application, first reported in 2018, and studies have assessed image quality, IGRT workflows, and photon/proton dosimetric validity. Most used paired training with registration, while results also examined anatomical-site generalisation and multi-region models.

III.B. CBCT-to-CT generation

  • CBCT-to-CT generation was the most recent CT-synthesis application, with its first paper published in 2018.128The review identifies this application as emerging later than the other CT-synthesis categories.Source: III.B. CBCT-to-CT generation · p. 19
  • Among 15 studies, five focused on CBCT image-quality improvement, whereas ten validated the transformation dosimetrically for photons, protons, or both.Dosimetric evaluations covered photons, protons, and combined photon-proton applications.Source: III.B. CBCT-to-CT generation · p. 19
  • Paired training predominated in 11 studies through rigid or deformable CBCT-to-CT registration, while only three studies used unpaired training.Eck et al.70 instead trained without registration, treating first-fraction CBCT as geometrically close to planning CT before later deformable registration for image-similarity analysis.Source: III.B. CBCT-to-CT generation · p. 19
  • A 2D cycle-GAN tested across head-and-neck, thorax, and pelvis achieved its best results in the pelvic region.The independently trained network was evaluated through image similarity, anatomical segmentation, and dosimetric accuracy within an IGART workflow.Source: III.B. CBCT-to-CT generation · p. 19
  • Maspero et al.88 found similar image-similarity results for three site-specific cycle-GANs and one model trained jointly across head-and-neck, breast, and lung.The comparison evaluated independently trained anatomical-site models against a single model trained on all sites together.Source: III.B. CBCT-to-CT generation · p. 19

III. RESULTS III.B. CBCT-to-CT generation

CBCT-to-CT studies evaluated supervised and unpaired GAN-based methods across pelvic, head-and-neck, and pancreatic radiotherapy applications. Results supported cross-scanner and cross-region transfer, GAN superiority over U-net baselines, and attention-gated cycle-GAN benefits for moving organs, while residual pancreatic discrepancies remained.

CBCT-to-CT generation

  • A 2.5D conditional GAN with feature matching trained on 135 pelvic patients achieved similar MAE on 15 different-scanner pelvic patients and ten head-and-neck patients, supporting transfer across scanners and regions.Zhang et al.89 compared GAN variants with U-net and found U-net statistically worse than every GAN configuration.Source: §III · p. 21
  • Among unpaired methods, Liang et al.132 found cycle-GAN outperformed DCGAN139 and PGGAN140 on both image similarity and image quality.Three works used cycle-GANs for unpaired training.Source: §III · p. 21
  • Most investigated anatomical regions were the head-and-neck and pelvis, with additional evaluation in pancreatic radiotherapy.The pancreatic study specifically addressed moving organs using attention gating.Source: §III · p. 21
  • For breath-hold stereotactic pancreatic radiotherapy, a 3D patch cycle-GAN with attention gate141 outperformed U-net and cycle-GAN without attention gating.Liu et al.135 reported statistically significant sCT improvement versus CBCT, although residual discrepancies remained at this anatomical site.Source: §III · p. 21

III.C. PET attenuation correction

For PET attenuation correction, deep-learning sCT methods comprise MR-to-CT and uncorrected PET-to-CT translations. Across these settings, reported results show improved PET accuracy, competitive performance with fewer alignment requirements, and applicability to heterogeneous or larger datasets.

PET attenuation correction

  • MR-to-CT methods dominate PET attenuation-correction studies, with 14 papers versus three using uncorrected PET-to-CT translation.MR-to-CT studies primarily used paired head-and-neck or pelvic data, whereas uncorrected PET-to-CT methods enabled full-body studies with larger cohorts.Source: III.C. PET attenuation correction · pp. 21, 24
  • DL prediction reduced corrected-PET SUV RMSE by factors of 4 for bone lesions and 1.5 for soft-tissue lesions compared with Dixon-based sCT.Other studies reported better DL-based attenuation correction than vendor atlas-based MRAC, while GAN augmentation was only slightly better than U-net without statistically relevant differences.Source: III.C. PET attenuation correction · p. 22
  • Cycle-GAN using unregistered MR/CT pairs outperformed atlas MRAC in DSC, MAE, and PETerr, with no significant difference from a registered-pair CNN.The authors argued that cycle-GAN can avoid perfectly aligned training data, although it requires more input data to improve output [150].Source: III.C. PET attenuation correction · p. 22
  • Pre-training a U-net on 19 healthy brains acquired with T1 GRE MRI led to faster training and slightly smaller residual error than random initialization of the U-net weights.Source: III.C. PET attenuation correction · p. 22
  • Uncorrected PET-to-CT methods achieved average PETerr below 1% despite global MAE around 110 HU versus 10–15 HU for site-specific MR-to-CT studies.These methods used up to 100 training and 39 testing patients, and results supported application across heterogeneous datasets and scanners.Source: III.C. PET attenuation correction · p. 24

IV. Discussion

This section discusses the clinical promise and feasibility of DL-based sCT generation across MR-only radiotherapy, CBCT-based adaptive radiotherapy, and PET attenuation correction, alongside limitations in image quality, generalisation, evaluation, and implementation. It also considers requirements for clinical adoption and promising extensions such as transfer learning, cross-modality translation, quality assurance, and uncertainty-based failure detection.

Deep learning considerations and trends

  • DL image translation can generate sCT in a few seconds, but training cohorts remain variable and limited by the availability of paired data.Reported training sets ranged from 7 to 242 patients, with 205 patients in one CBCT study; data augmentation is commonly used.Source: §IV · p. 24, Deep learning considerations and trends · p. 27

IV. DISCUSSION

  • MR-only RT is the largest category, with 51 papers showing high image similarity and dose differences < 1% for photon and proton therapy in pelvis and brain/H&N.Motion-affected anatomy shows promising feasibility, but no DL-based 4D sCT study was identified; pediatric MR-only RT is an important application.Source: §IV · pp. 24–25
  • Clinical adoption requires further geometric and magnetic-field dosimetric validation, especially for position verification, high-field MR-guided RT, and additional anatomical sites.Only Gupta et al.110 and Olberg et al.125 investigated alignment accuracy for brain and breast, while high-field MR-guided RT achieved dose differences < 1% for breast cases.Source: §IV · p. 25
  • CBCT-to-CT improves image quality and speed, enables accurate photon and proton dose calculations, and may support image guidance and organ-at-risk segmentation for adaptive RT.Clinical translation remains constrained by CBCT–CT field-of-view differences, which can hinder replanning unless missing anatomy is addressed.Source: §IV · p. 26
  • Direct PET-to-μ-map prediction is an auspicious future direction, while it remains unclear whether CBCT quality enhancement is better formulated as synthesis or reconstruction.Initial reconstruction-network studies improved generalisation to other anatomy, but this research remains outside the review’s image-to-CT scope.Source: §IV · pp. 26–27
  • DL-based sCT consistently outperformed commercial MRAC methods and may overcome most current PET attenuation-correction limitations.The review considered sCT from MRI or uncorrected PET, while direct μ-map prediction from uncorrected PET was outside its image-to-CT scope.Source: §IV · p. 27

Deep learning considerations and trends

  • Training with up to fifty patients increases image similarity, but the optimal dataset size depends on anatomy and motion; imbalanced data may promote overfitting.The review emphasizes investigating training-set size and balancing datasets to identify the minimum amount needed for state-of-the-art performance.Source: Deep learning considerations and trends · pp. 27–28
  • GANs were most popular, but comparisons reported statistically better, similar, or worse performance than U-net or other CNNs, implicating loss functions as an underexplored determinant.Loss functions drive network learning but have received less investigation than architecture.Source: Deep learning considerations and trends · p. 28
  • Comparing paired and unpaired training, paired training led to slightly better performances, although the differences were not always statistically significant71,80,95.Source: Deep learning considerations and trends · p. 28
  • Studies mainly addressed brain, head-and-neck, and pelvic regions; fewer studies examined thorax and abdomen, which represent a more challenging patient population because of organ motion.In MR-only radiotherapy, results regarding the best-performing spatial approach were contradictory, while 2D+ increased sCT quality compared with a single 2D view.Source: Deep learning considerations and trends · p. 28
  • Attention gates significantly improved predictions in a cycle-GAN for pancreas adaptive radiotherapy, and attention may support future multitask systems that derive sCT while delineating structures.The review identifies attention mechanisms as a developing direction for focusing on target structures and enabling multi-task applications.Source: Deep learning considerations and trends · p. 29
  • No reviewed method was clinically implemented; adoption is constrained by regulatory requirements, uncertain position-verification impact in categories I/II, heterogeneous metrics, and the absence of public grand-challenge datasets.The review advocates consensus-based reporting requirements and broader multicentre evaluation, noting that existing multicentre studies cover only MR-only radiotherapy.Source: Deep learning considerations and trends · pp. 29–30
  • Transfer learning could support generalisation across MRI contrasts, CBCT protocols, and anatomies, but no sCT study had investigated this by December 2020.Li et al.216 showed transfer learning could facilitate training across different MRI contrasts.Source: Deep learning considerations and trends · p. 31
  • Software-based QA, potentially using phantoms or CNN-generated quality metrics, is needed because users cannot reliably judge sCT quality unless it is inferior.Phantom manufacturing is particularly challenging for MR-based sCT because phantoms must reproduce MRI and CT contrast, although additive manufacturing shows potential.Source: Deep learning considerations and trends · p. 31
  • Uncertainty estimation remains only superficially explored for sCT generation, motivating automatic failure-case detection so inaccurate sCTs can trigger CT rescanning or manual adjustment.Existing approaches use multi-task Bayesian modelling, ensembles, or dropout-based variational inference, but criteria for identifying failures remain a future need.Source: Deep learning considerations and trends · p. 31

Beyond sCT for radiotherapy

  • Beyond radiotherapy sCT generation, related applications include synthetic dual-energy CT, lesion detection, surgical planning, multimodal registration, MR-guided focused ultrasound, and cross-modality segmentation.Reviewed sCT methods can also translate other image modalities, including MRI-to-CT and CT-to-MRI generation for improving segmentation robustness or exploiting annotations from another modality.Source: Beyond sCT for radiotherapy · pp. 31–32

V. Conclusion

The review covers DL-based sCT generation for MR-based treatment planning, CBCT-based adaptive radiotherapy, and PET attenuation mapping. Across these applications, sCT appears feasible in several anatomical sites, but clinical deployment requires broader validation and commissioning.

V. Conclusion

  • The review compared imaging protocols, DL architectures, and performance using the most popular reported metrics across three clinical categories.The categories were MR-based CT replacement, CBCT-based adaptive radiotherapy, and PET attenuation-map generation.Source: §V · p. 32
  • DL-based sCT generation is an active and growing research area with an extensive body of work and a bright future.The conclusion frames continued research as central to advancing these methods.Source: §V · p. 32
  • Dose differences from CT-based planning were < 1% for several anatomical sites, including H&N/brain and pelvis, while PET attenuation correction outperformed standard MRAC methods.These findings support the feasibility of DL-based sCT generation in radiotherapy and PET.Source: §V · p. 32
  • Clinical translation requires evaluating generalisation across multiple centres and establishing comprehensive commissioning and QA methods to ensure treatment efficacy and patient safety.These steps are necessary before DL-based sCT techniques can be widely adopted in clinical practice.Source: §V · p. 32

Appendix · VIII. Acronyms and abbreviations

The appendix documents the review’s database search strategy and provides an acronym glossary for methods, imaging modalities, evaluation metrics, and clinical applications. The search combined radiotherapy, imaging, PET, synthetic-CT, and deep-learning terminology across PubMed, Scopus, and Web of Science, with inclusion and exclusion criteria summarized in Figure 4.

Appendix

  • The review searched PubMed, Scopus, and Web of Science using Title/Abstract/Keywords fields.Source: Appendix · p. 33
  • The query combined radiotherapy, imaging, CBCT, PET, attenuation-correction, and attenuation-map terms with synthetic-CT variants.Source: Appendix · p. 33
  • Deep-learning query terms included deep learning, convolutional network, CNN, GAN, GANN, and artificial intelligence.Source: Appendix · p. 33
  • Figure 4 schematizes the review’s search inclusion and exclusion criteria, including time window, keywords, article type, content, and three categories.Source: Figure 4 · p. 34

VIII. Acronyms and abbreviations

  • The acronym glossary defines core imaging, learning, anatomical, and radiotherapy terms, including CBCT, CNNs, CT, DL, GANs, H&N, IGART, MR, PET, and sCT.Source: §VIII · p. 34
  • It also defines quantitative evaluation and reconstruction measures, including MAE, MSE, RMSE, PSNR, SSIM, DSC, FID, GPR, DPR, PETerr, and PET|err|.Source: §VIII · pp. 34–35

VIII. ACRONYMS AND ABBREVIATIONS

  • Additional entries cover acquisition, geometry, anatomy, and PET terminology, including FLAIR, GRE, mUTE, UTE, ZTE, FOV, ROI, OARS, SUV, VOI, and µ −map.Source: §VIII · pp. 34–35

Related papers

Loading 2102.02734v2…