Dataset Information

Deep-Learning-Based Neural Tissue Segmentation of MRI in Multiple Sclerosis: Effect of Training Set Size.

ABSTRACT:

Background

The dependence of deep-learning (DL)-based segmentation accuracy of brain MRI on the training size is not known.

Purpose

To determine the required training size for a desired accuracy in brain MRI segmentation in multiple sclerosis (MS) using DL.

Study type

Retrospective analysis of MRI data acquired as part of a multicenter clinical trial.

Study population

In all, 1008 patients with clinically definite MS.

Field strength/sequence

MRIs were acquired at 1.5T and 3T scanners manufactured by GE, Philips, and Siemens with dual turbo spin echo, FLAIR, and T₁ -weighted turbo spin echo sequences.

Assessment

Segmentation results using an automated analysis pipeline and validated by two neuroimaging experts served as the ground truth. A DL model, based on a fully convolutional neural network, was trained separately using 16 different training sizes. The segmentation accuracy as a function of the training size was determined. These data were fitted to the learning curve for estimating the required training size for desired accuracy.

Statistical tests

The performance of the network was evaluated by calculating the Dice similarity coefficient (DSC), and lesion true-positive and false-positive rates.

Results

The DSC for lesions showed much stronger dependency on the sample size than gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF). When the training size was increased from 10 to 800 the DSC values varied from 0.00 to 0.86 ± 0.016 for T₂ lesions, 0.87 ± 009 to 0.94 ± 0.004 for GM, 0.86 ± 0.08 to 0.94 ± 0.005 for WM, and 0.91 ± 0.009 to 0.96 ± 0.003 for CSF.

Data conclusion

Excellent segmentation was achieved with a training size as small as 10 image volumes for GM, WM, and CSF. In contrast, a training size of at least 50 image volumes was necessary for adequate lesion segmentation.

Level of evidence

1 Technical Efficacy Stage: 1 J. Magn. Reson. Imaging 2020;51:1487-1496.

SUBMITTER: Narayana PA

PROVIDER: S-EPMC7165037 | biostudies-literature |

REPOSITORIES: biostudies-literature

ACCESS DATA

Similar Datasets

Project description:BackgroundProstate segmentation is an essential step in computer-aided detection and diagnosis systems for prostate cancer. Deep learning (DL)-based methods provide good performance for prostate gland and zones segmentation, but little is known about the impact of manual segmentation (that is, label) selection on their performance. In this work, we investigated these effects by obtaining two different expert label-sets for the PROSTATEx I challenge training dataset (n = 198) and using them, in addition to an in-house dataset (n = 233), to assess the effect on segmentation performance. The automatic segmentation method we used was nnU-Net.ResultsThe selection of training/testing label-set had a significant (p < 0.001) impact on model performance. Furthermore, it was found that model performance was significantly (p < 0.001) higher when the model was trained and tested with the same label-set. Moreover, the results showed that agreement between automatic segmentations was significantly (p < 0.0001) higher than agreement between manual segmentations and that the models were able to outperform the human label-sets used to train them.ConclusionsWe investigated the impact of label-set selection on the performance of a DL-based prostate segmentation model. We found that the use of different sets of manual prostate gland and zone segmentations has a measurable impact on model performance. Nevertheless, DL-based segmentation appeared to have a greater inter-reader agreement than manual segmentation. More thought should be given to the label-set, with a focus on multicenter manual segmentation and agreement on common procedures.Critical relevance statementLabel-set selection significantly impacts the performance of a deep learning-based prostate segmentation model. Models using different label-set showed higher agreement than manual segmentations.Key points• Label-set selection has a significant impact on the performance of automatic segmentation models. • Deep learning-based models demonstrated true learning rather than simply mimicking the label-set. • Automatic segmentation appears to have a greater inter-reader agreement than manual segmentation.

Project description:PurposeTo improve the robustness of deep learning-based glioblastoma segmentation in a clinical setting with sparsified datasets.Materials and methodsIn this retrospective study, preoperative T1-weighted, T2-weighted, T2-weighted fluid-attenuated inversion recovery, and postcontrast T1-weighted MRI from 117 patients (median age, 64 years; interquartile range [IQR], 55-73 years; 76 men) included within the Multimodal Brain Tumor Image Segmentation (BraTS) dataset plus a clinical dataset (2012-2013) with similar imaging modalities of 634 patients (median age, 59 years; IQR, 49-69 years; 382 men) with glioblastoma from six hospitals were used. Expert tumor delineations on the postcontrast images were available, but for various clinical datasets, one or more sequences were missing. The convolutional neural network, DeepMedic, was trained on combinations of complete and incomplete data with and without site-specific data. Sparsified training was introduced, which randomly simulated missing sequences during training. The effects of sparsified training and center-specific training were tested using Wilcoxon signed rank tests for paired measurements.ResultsA model trained exclusively on BraTS data reached a median Dice score of 0.81 for segmentation on BraTS test data but only 0.49 on the clinical data. Sparsified training improved performance (adjusted P < .05), even when excluding test data with missing sequences, to median Dice score of 0.67. Inclusion of site-specific data during sparsified training led to higher model performance Dice scores greater than 0.8, on par with a model based on all complete and incomplete data. For the model using BraTS and clinical training data, inclusion of site-specific data or sparsified training was of no consequence.ConclusionAccurate and automatic segmentation of glioblastoma on clinical scans is feasible using a model based on large, heterogeneous, and partially incomplete datasets. Sparsified training may boost the performance of a smaller model based on public and site-specific data.Supplemental material is available for this article.Published under a CC BY 4.0 license.