Dataset Information

Learning deep abdominal CT registration through adaptive loss weighting and synthetic data generation.

ABSTRACT:

Purpose

This study aims to explore training strategies to improve convolutional neural network-based image-to-image deformable registration for abdominal imaging.

Methods

Different training strategies, loss functions, and transfer learning schemes were considered. Furthermore, an augmentation layer which generates artificial training image pairs on-the-fly was proposed, in addition to a loss layer that enables dynamic loss weighting.

Results

Guiding registration using segmentations in the training step proved beneficial for deep-learning-based image registration. Finetuning the pretrained model from the brain MRI dataset to the abdominal CT dataset further improved performance on the latter application, removing the need for a large dataset to yield satisfactory performance. Dynamic loss weighting also marginally improved performance, all without impacting inference runtime.

Conclusion

Using simple concepts, we improved the performance of a commonly used deep image registration architecture, VoxelMorph. In future work, our framework, DDMR, should be validated on different datasets to further assess its value.

SUBMITTER: Perez de Frutos J

PROVIDER: S-EPMC9956065 | biostudies-literature | 2023

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Learning deep abdominal CT registration through adaptive loss weighting and synthetic data generation.

Pérez de Frutos Javier J Pedersen André A Pelanis Egidijus E Bouget David D Survarachakan Shanmugapriya S Langø Thomas T Elle Ole-Jakob OJ Lindseth Frank F

PloS one 20230224 2

<h4>Purpose</h4>This study aims to explore training strategies to improve convolutional neural network-based image-to-image deformable registration for abdominal imaging.<h4>Methods</h4>Different training strategies, loss functions, and transfer learning schemes were considered. Furthermore, an augmentation layer which generates artificial training image pairs on-the-fly was proposed, in addition to a loss layer that enables dynamic loss weighting.<h4>Results</h4>Guiding registration using segme ...[more]

PMID: 36827289

Similar Datasets

Project description:BackgroundMost studies on synthetic computed tomography (sCT) generation for brain rely on in-house developed methods. They often focus on performance rather than clinical feasibility. Therefore, the aim of this work was to validate sCT images generated using a commercially available software, based on a convolutional neural network (CNN) algorithm, to enable MRI-only treatment planning for the brain in a clinical setting.MethodsThis prospective study included 20 patients with brain malignancies of which 14 had areas of resected skull bone due to surgery. A Dixon magnetic resonance (MR) acquisition sequence for sCT generation was added to the clinical brain MR-protocol. The corresponding sCT images were provided by the software MRI Planner (Spectronic Medical AB, Sweden). sCT images were rigidly registered and resampled to CT for each patient. Treatment plans were optimized on CT and recalculated on sCT images for evaluation of dosimetric and geometric endpoints. Further analysis was also performed for the post-surgical cases. Clinical robustness in patient setup verification was assessed by rigidly registering cone beam CT (CBCT) to sCT and CT images, respectively.ResultsAll sCT images were successfully generated. Areas of bone resection due to surgery were accurately depicted. Mean absolute error of the sCT images within the body contour for all patients was 62.2 ± 4.1 HU. Average absorbed dose differences were below 0.2% for parameters evaluated for both targets and organs at risk. Mean pass rate of global gamma (1%/1 mm) for all patients was 100.0 ± 0.0% within PTV and 99.1 ± 0.6% for the full dose distribution. No clinically relevant deviations were found in the CBCT-sCT vs CBCT-CT image registrations. In addition, mean values of voxel-wise patient specific geometric distortion in the Dixon images for sCT generation were below 0.1 mm for soft tissue, and below 0.2 mm for air and bone.ConclusionsThis work successfully validated a commercially available CNN-based software for sCT generation. Results were comparable for sCT and CT images in both dosimetric and geometric evaluation, for both patients with and without anatomical anomalies. Thus, MRI Planner is feasible to use for radiotherapy treatment planning of brain tumours.

Project description:Background Assessment of liver lesions is constrained as CT radiation doses are lowered; evidence suggests deep learning reconstructions mitigate such effects. Purpose To evaluate liver metastases and image quality between reduced-dose deep learning image reconstruction (DLIR) and standard-dose filtered back projection (FBP) contrast-enhanced abdominal CT. Materials and Methods In this prospective Health Insurance Portability and Accountability Act-compliant study (September 2019 through April 2021), participants with biopsy-proven colorectal cancer and liver metastases at baseline CT underwent standard-dose and reduced-dose portal venous abdominal CT in the same breath hold. Three radiologists detected and characterized lesions at standard-dose FBP and reduced-dose DLIR, reported confidence, and scored image quality. Contrast-to-noise ratios for liver metastases were recorded. Summary statistics were reported, and a generalized linear mixed model was used. Results Fifty-one participants (mean age ± standard deviation, 57 years ± 13; 31 men) were evaluated. The mean volume CT dose index was 65.1% lower with reduced-dose CT (12.2 mGy) than with standard-dose CT (34.9 mGy). A total of 161 lesions (127 metastases, 34 benign lesions) with a mean size of 0.7 cm ± 0.3 were identified. Subjective image quality of reduced-dose DLIR was superior to that of standard-dose FBP (P < .001). The mean contrast-to-noise ratio for liver metastases of reduced-dose DLIR (3.9 ± 1.7) was higher than that of standard-dose FBP (3.5 ± 1.4) (P < .001). Differences in detection were identified only for lesions 0.5 cm or smaller: 63 of 65 lesions detected with standard-dose FBP (96.9%; 95% CI: 89.3, 99.6) and 47 lesions with reduced-dose DLIR (72.3%; 95% CI: 59.8, 82.7). Lesion accuracy with standard-dose FBP and reduced-dose DLIR was 80.1% (95% CI: 73.1, 86.0; 129 of 161 lesions) and 67.1% (95% CI: 59.3, 74.3; 108 of 161 lesions), respectively (P = .01). Lower lesion confidence was reported with a reduced dose (P < .001). Conclusion Deep learning image reconstruction (DLIR) improved CT image quality at 65% radiation dose reduction while preserving detection of liver lesions larger than 0.5 cm. Reduced-dose DLIR demonstrated overall inferior characterization of liver lesions and reader confidence. Clinical trial registration no. NCT03151564 © RSNA, 2022 Online supplemental material is available for this article.

Project description:PurposeTo automate skeletal muscle segmentation in a pediatric population using convolutional neural networks that identify and segment the L3 level at CT.Materials and methodsIn this retrospective study, two sets of U-Net-based models were developed to identify the L3 level in the sagittal plane and segment the skeletal muscle from the corresponding axial image. For model development, 370 patients (sampled uniformly across age group from 0 to 18 years and including both sexes) were selected between January 2009 and January 2019, and ground truth L3 location and skeletal muscle segmentation were manually defined. Twenty percent (74 of 370) of the examinations were reserved for testing the L3 locator and muscle segmentation, while the remaining were used for training. For the L3 locator models, maximum intensity projections (MIPs) from a fixed number of central sections of sagittal reformats (either 12 or 18 sections) were used as input with or without transfer learning using an L3 localizer trained on an external dataset (four models total). For the skeletal muscle segmentation models, two loss functions (weighted Dice similarity coefficient [DSC] and binary cross-entropy) were used on models trained with or without data augmentation (four models total). Outputs from each model were compared with ground truth, and the mean relative error and DSC from each of the models were compared with one another.ResultsL3 section detection trained with an 18-section MIP model with transfer learning had a mean error of 3.23 mm ± 2.61 standard deviation, which was within the reconstructed image thickness (3 or 5 mm). Skeletal muscle segmentation trained with the weighted DSC loss model without data augmentation had a mean DSC of 0.93 ± 0.03 and mean relative error of 0.04 ± 0.04.ConclusionConvolutional neural network models accurately identified the L3 level and segmented the skeletal muscle on pediatric CT scans.Supplemental material is available for this article.See also the commentary by Cadrin-Chênevert in this issue.© RSNA, 2021.

Dataset Information

Learning deep abdominal CT registration through adaptive loss weighting and synthetic data generation.

Purpose

Methods

Results

Conclusion

Publications

Learning deep abdominal CT registration through adaptive loss weighting and synthetic data generation.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets