Comparison of 2D vs. 3D U-Net Organ Segmentation in abdominal 3D CT images

Zettler, Nico; Mastmeyer, André

doi:10.48550/arxiv.2107.04062

Cited by 7 publications

(4 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A two-dimensional (2D) MultiResUNet implementation was used, such that individual image slices are used as model input. Upon reviewing the literature to compare the performance of OAR segmentation models using 2D and three-dimensional (3D) U-Net [23,24], it was found that 2D U-Net often performs as well or better than 3D U-Net, and we, therefore, hypothesized that repeating model training with a 3D-architecture would not be significantly beneficial for the present study.…”

Section: Modelmentioning

confidence: 99%

Development of a CT-Based Auto-Segmentation Model for Prostate-Specific Membrane Antigen (PSMA) Positron Emission Tomography-Delineated Tubarial Glands

Caleb¹,

Jung²,

Rahmim³

et al. 2022

Cureus

View full text Add to dashboard Cite

Modern inverse planning radiotherapy techniques allow for organs at risk (OARs) to evade radiation doses that they would have been subjected to with earlier techniques. The extent to which patient outcomes may be improved using these techniques depends on the delineation accuracy of target volumes and OARs on medical images as well as clinical dose constraints applied to regions of interest (ROIs). The recent discovery of bilateral "tubarial" salivary glands, which were found in the nasopharynx using prostate-specific membrane antigen (PSMA) positron emission tomography (PET), raises concerns over how dose to this region might affect patient outcomes. The dose response of the major salivary glands is known to be variable, and it is possible that the dose in tubarial glands constitutes a missing variable in the optimization of head and neck (HN) radiotherapy plans. A necessary first step toward conducting clinical studies that include the tubarial glands in plan optimization is to develop methods for delineating these glands without the use of PSMA PET images, as their acquisition is not considered a part of the standard of care for HN patients. In this study, we develop an open-source program, Organogenesis, for the auto-segmentation of tubarial glands using only computed tomography (CT) images. A predictive model is trained using contours derived from PSMA PET images, allowing for accurate delineation of tubarial glands, which cannot be manually contoured using CT only. Organogenesis provides a predictive model for tubarial glands that can be iteratively improved on with additional data, creating a viable pathway to clinical studies that can assess the importance of incorporating tubarial glands into HN radiotherapy plan optimization.

show abstract

Section: Modelmentioning

confidence: 99%

Development of a CT-Based Auto-Segmentation Model for Prostate-Specific Membrane Antigen (PSMA) Positron Emission Tomography-Delineated Tubarial Glands

Caleb¹,

Jung²,

Rahmim³

et al. 2022

Cureus

View full text Add to dashboard Cite

show abstract

“…[46][47][48] Because of this issue, studies have shown that 2D architectures can sometimes be trained more optimally and perform equivalently or even outperform 3D networks. [46][47][48] Our network was applied to the segmentation of the prostate gland only primarily because our primary CT dataset was limited to the availability of the prostate contours only, and our secondary dataset consisted of multiple clinician contours for the prostate gland only.We are working on the acquisition of manual contours for normal organs, which we will evaluate as part of a future study. Additionally, we intend to compare results of our network with that of other newer architectures, such as deeply supervised U-Net, spatial transformer network, and mask R-CNN.…”

Section: Discussionmentioning

confidence: 99%

“…However, 3D models are typically much more complex and require significantly more computational time, and memory resources, which confound the optimal training of these models. [46][47][48] Because of this issue, studies have shown that 2D architectures can sometimes be trained more optimally and perform equivalently or even outperform 3D networks. [46][47][48] Our network was applied to the segmentation of the prostate gland only primarily because our primary CT dataset was limited to the availability of the prostate contours only, and our secondary dataset consisted of multiple clinician contours for the prostate gland only.We are working on the acquisition of manual contours for normal organs, which we will evaluate as part of a future study.…”

Section: Discussionmentioning

confidence: 99%

An uncertainty‐aware deep learning architecture with outlier mitigation for prostate gland segmentation in radiotherapy treatment planning

Bagher-Ebadian

Gardner

et al. 2022

Medical Physics

View full text Add to dashboard Cite

Purpose Task automation is essential for efficient and consistent image segmentation in radiation oncology. We report on a deep learning architecture, comprising a U‐Net and a variational autoencoder (VAE) for automatic contouring of the prostate gland incorporating interobserver variation for radiotherapy treatment planning. The U‐Net/VAE generates an ensemble set of segmentations for each image CT slice. A novel outlier mitigation (OM) technique was implemented to enhance the model segmentation accuracy. Methods The primary source dataset (source_prim) consisted of 19 200 CT slices (from 300 patient planning CT image datasets) with manually contoured prostate glands. A smaller secondary source dataset (source_sec) comprised 640 CT slices (from 10 patient CT datasets), where prostate glands were segmented by 5 independent physicians on each dataset to account for interobserver variability. Data augmentation via random rotation (<5 degrees), cropping, and horizontal flipping was applied to each dataset to increase sample size by a factor of 100. A probabilistic hierarchical U‐Net with VAE was implemented and pretrained using the augmented source_prim dataset for 30 epochs. Model parameters of the U‐Net/VAE were fine‐tuned using the augmented source_sec dataset for 100 epochs. After the first round of training, outlier contours in the training dataset were automatically detected and replaced by the most accurate contours (based on Dice similarity coefficient, DSC) generated by the model. The U‐Net/OM‐VAE was retrained using the revised training dataset. Metrics for comparison included DSC, Hausdorff distance (HD, mm), normalized cross‐correlation (NCC) coefficient, and center‐of‐mass (COM) distance (mm). Results Results for U‐Net/OM‐VAE with outliers replaced in the training dataset versus U‐Net/VAE without OM were as follows: DSC = 0.82 ± 0.01 versus 0.80 ± 0.02 (p = 0.019), HD = 9.18 ± 1.22 versus 10.18 ± 1.35 mm (p = 0.043), NCC = 0.59 ± 0.07 versus 0.62 ± 0.06, and COM = 3.36 ± 0.81 versus 4.77 ± 0.96 mm over the average of 15 contours. For the average of 15 highest accuracy contours, values were as follows: DSC = 0.90 ± 0.02 versus 0.85 ± 0.02, HD = 5.47 ± 0.02 versus 7.54 ± 1.36 mm, and COM = 1.03 ± 0.58 versus 1.46 ± 0.68 mm (p < 0.03 for all metrics). Results for the U‐Net/OM‐VAE with outliers removed were as follows: DSC = 0.78 ± 0.01, HD = 10.65 ± 1.95 mm, NCC = 0.46 ± 0.10, COM = 4.17 ± 0.79 mm for the average of 15 contours, and DSC = 0.88 ± 0.02, HD = 7.00 ± 1.17 mm, COM = 1.58 ± 0.63 mm for the average of 15 highest accuracy contours. All metrics for U‐Net/VAE trained on the source_prim and source_sec datasets via pretraining, followed by fine‐tuning, show statistically significant improvement over that trained on the source_sec dataset only. Finally, all metrics for U‐Net/VAE with or without OM showed statistically significant improvement over those for the standard U‐Net. Conclusions A VAE combined with a hierarchical U‐Net and an OM strategy (U‐Net/OM‐VAE) demonstrates promise toward capt...

show abstract

“…Numerous approaches and data types have been used in considerable research on the effectiveness of models in a variety of activities. Zettler et al show that 2D U-Net models are more effective than 3D U-Net models in terms of speed and low memory costs [3]. Even while the 3D model had slightly more favorable results than the other model, they all concluded that this couldn’t possibly justify the computer resources that were wasted on it.…”

Section: Related Workmentioning

confidence: 99%

Unlocking Robotic Potential Through Modern Organ Segmentation

Chaudhary

2023

Preprint

View full text Add to dashboard Cite

Deep learning has revolutionized the approach to complex data-driven problems, specifically in medical imaging, where its techniques have significantly raised efficiency in organ segmentation. The urgent need to enhance the depth and precision of organ-based classification is an essential step towards automation of medical operation and diagnostics. The research aims to investigate the effect and potential advantages transformer models have on binary semantic segmentation, the method utilized for the project. Hence, I employed the SegFormer model, for its lightweight architecture, as the primary deep learning model, alongside the Unet. A custom 2D computerized tomography (CT) scan dataset was assembled, CT-Org2D through meticulous operations. Extensive experiments showed that, in contrast to the selected models, the task’s simplicity required a redesigned Unet architecture with reduced complexity. This model yielded impressive results: Precision, Recall, and IOU scores of 0.91, 0.92, and 0.85 respectively. The research serves as a starting point, motivating further exploration, through different methodologies, to achieve even greater efficiency in organ segmentation.

show abstract

Comparison of 2D vs. 3D U-Net Organ Segmentation in abdominal 3D CT images

Cited by 7 publications

References 0 publications

Development of a CT-Based Auto-Segmentation Model for Prostate-Specific Membrane Antigen (PSMA) Positron Emission Tomography-Delineated Tubarial Glands

Development of a CT-Based Auto-Segmentation Model for Prostate-Specific Membrane Antigen (PSMA) Positron Emission Tomography-Delineated Tubarial Glands

An uncertainty‐aware deep learning architecture with outlier mitigation for prostate gland segmentation in radiotherapy treatment planning

Unlocking Robotic Potential Through Modern Organ Segmentation

Contact Info

Product

Resources

About