Dataset Information

An Automatic Deep Learning-Based Workflow for Glioblastoma Survival Prediction Using Preoperative Multimodal MR Images: A Feasibility Study.

ABSTRACT:

Purpose

Most radiomic studies use the features extracted from the manually drawn tumor contours for classification or survival prediction. However, large interobserver segmentation variations lead to inconsistent features and hence introduce more challenges in constructing robust prediction models. Here, we proposed an automatic workflow for glioblastoma (GBM) survival prediction based on multimodal magnetic resonance (MR) images.

Methods and materials

Two hundred eighty-five patients with glioma (210 GBM, 75 low-grade glioma) were included. One hundred sixty-three of the patients with GBM had overall survival data. Every patient had 4 preoperative MR images and manually drawn tumor contours. A 3-dimensional convolutional neural network, VGG-Seg, was trained and validated using 122 patients with glioma for automatic GBM segmentation. The trained VGG-Seg was applied to the remaining 163 patients with GBM to generate their autosegmented tumor contours. The handcrafted and deep learning (DL)-based radiomic features were extracted from the autosegmented contours using explicitly designed algorithms and a pretrained convolutional neural network, respectively. One hundred sixty-three patients with GBM were randomly split into training (n = 122) and testing (n = 41) sets for survival analysis. Cox regression models were trained to construct the handcrafted and DL-based signatures. The prognostic powers of the 2 signatures were evaluated and compared.

Results

The VGG-Seg achieved a mean Dice coefficient of 0.86 across 163 patients with GBM for GBM segmentation. The handcrafted signature achieved a C-index of 0.64 (95% confidence interval, 0.55-0.73), whereas the DL-based signature achieved a C-index of 0.67 (95% confidence interval, 0.57-0.77). Unlike the handcrafted signature, the DL-based signature successfully stratified testing patients into 2 prognostically distinct groups.

Conclusions

The VGG-Seg generated accurate GBM contours from 4 MR images. The DL-based signature achieved a numerically higher C-index than the handcrafted signature and significant patient stratification. The proposed automatic workflow demonstrated the potential of improving patient stratification and survival prediction in patients with GBM.

SUBMITTER: Fu J

PROVIDER: S-EPMC8377554 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:PURPOSE:To automate the estimation of swallowing motion from 2D MR cine images using deformable registration for future applications of personalized margin reduction in head and neck radiotherapy and outcome assessment of radiation-associated dysphagia. METHODS:Twenty-one patients with serial 2D FSPGR-MR cine scans of the head and neck conducted through the course of definitive radiotherapy for oropharyngeal cancer. Included patients had at least one cine scan before, during, or after radiotherapy, with a total of 52 cine scans. Contours of 7 swallowing related regions-of-interest (ROIs), including pharyngeal constrictor, epiglottis, base of tongue, geniohyoid, hyoid, soft palate, and larynx, were manually delineated from consecutive frames of the cine scan covering at least one swallowing cycle. We applied a modified thin-plate-spline robust-point-matching algorithm to register the point sets of each ROI automatically over frames. The deformation vector fields from the registration were then used to estimate the motion during swallowing for each ROI. Registration errors were estimated by comparing the deformed contours with the manual contours. RESULTS:On average 22 frames of each cine scan were contoured. The registration for one cine scan (7 ROIs over 22 frames) on average took roughly 22 minutes. A number of 8018 registrations were successfully batch processed without human interaction after the contours were drawn. The average registration error for all ROIs and all patients was 0.36 mm (range: 0.06 mm- 2.06 mm). Larynx had the average largest motion in superior direction of all structures under consideration (range: 0.0 mm- 58.7 mm). Geniohyoid had the smallest overall motion of all ROIs under consideration and the superior-inferior motion was larger than the anterior-posterior motion for all ROIs. CONCLUSION:We developed and validated a deformable registration framework to automate the estimation of swallowing motion from 2D MR cine scans.

Project description:PurposeTo develop and evaluate a complete deep learning pipeline that allows fully automated end-diastolic left ventricle (LV) cardiac MRI segmentation, including trabeculations and automatic quality control of the predicted segmentation.Materials and methodsThis multicenter retrospective study includes training, validation, and testing datasets of 272, 27, and 150 cardiac MR images, respectively, collected between 2012 and 2018. The reference standard was the manual segmentation of four LV anatomic structures performed on end-diastolic short-axis cine cardiac MRI: LV trabeculations, LV myocardium, LV papillary muscles, and the LV blood cavity. The automatic pipeline was composed of five steps with a DenseNet architecture. Intraobserver agreement, interobserver agreement, and interaction time were recorded. The analysis includes the correlation between the manual and automated segmentation, a reproducibility comparison, and Bland-Altman plots.ResultsThe automated method achieved mean Dice coefficients of 0.96 ± 0.01 (standard deviation) for LV blood cavity, 0.89 ± 0.03 for LV myocardium, and 0.62 ± 0.08 for LV trabeculation (mean absolute error, 3.63 g ± 3.4). Automatic quantification of LV end-diastolic volume, LV myocardium mass, LV trabeculation, and trabeculation mass-to-total myocardial mass (TMM) ratio showed a significant correlation with the manual measures (r = 0.99, 0.99, 0.90, and 0.83, respectively; all P < .01). On a subset of 48 patients, the mean Dice value for LV trabeculation was 0.63 ± 0.10 or higher compared with the human interobserver (0.44 ± 0.09; P < .01) and intraobserver measures (0.58 ± 0.09; P < .01). Automatic quantification of the trabeculation mass-to-TMM ratio had a higher correlation (0.92) compared with the intra- and interobserver measures (0.74 and 0.39, respectively; both P < .01).ConclusionAutomated deep learning framework can achieve reproducible and quality-controlled segmentation of cardiac trabeculations, outperforming inter- and intraobserver analyses.Supplemental material is available for this article.© RSNA, 2020.