An open-source nnU-net algorithm for automatic segmentation of MRI scans in the male pelvis for adaptive radiotherapy
Ebbe Laugaard LorenzenBahar ÇelikNis SarupLars DysagerR.L. ChristiansenA. BertelsenUffe BernchouS. AgergaardMaximilian KonradC. BrinkFaisal MahmoodTine SchytteChristina Junker Nyborg
3
Citation
13
Reference
10
Related Paper
Citation Trend
Abstract:
Adaptive MRI-guided radiotherapy (MRIgRT) requires accurate and efficient segmentation of organs and targets on MRI scans. Manual segmentation is time-consuming and variable, while deformable image registration (DIR)-based contour propagation may not account for large anatomical changes. Therefore, we developed and evaluated an automatic segmentation method using the nnU-net framework.The network was trained on 38 patients (76 scans) with localized prostate cancer and tested on 30 patients (60 scans) with localized prostate, metastatic prostate, or bladder cancer treated at a 1.5 T MRI-linac at our institution. The performance of the network was compared with the current clinical workflow based on DIR. The segmentation accuracy was evaluated using the Dice similarity coefficient (DSC), mean surface distance (MSD), and Hausdorff distance (HD) metrics.The trained network successfully segmented all 600 structures in the test set. High similarity was obtained for most structures, with 90% of the contours having a DSC above 0.9 and 86% having an MSD below 1 mm. The largest discrepancies were found in the sigmoid and colon structures. Stratified analysis on cancer type showed that the best performance was seen in the same type of patients that the model was trained on (localized prostate). Especially in patients with bladder cancer, the performance was lower for the bladder and the surrounding organs. A complete automatic delineation workflow took approximately 1 minute. Compared with contour transfer based on the clinically used DIR algorithm, the nnU-net performed statistically better across all organs, with the most significant gain in using the nnU-net seen for organs subject to more considerable volumetric changes due to variation in the filling of the rectum, bladder, bowel, and sigmoid.We successfully trained and tested a network for automatically segmenting organs and targets for MRIgRT in the male pelvis region. Good test results were seen for the trained nnU-net, with test results outperforming the current clinical practice using DIR-based contour propagation at the 1.5 T MRI-linac. The trained network is sufficiently fast and accurate for clinical use in an online setting for MRIgRT. The model is provided as open-source.Keywords:
Hausdorff distance
Sørensen–Dice coefficient
Tissue phase mapping (TPM) is an MRI technique for quantification of regional biventricular myocardial velocities. Despite its potential, clinical use is limited due to the requisite labor‐intensive manual segmentation of cardiac contours for all time frames. The purpose of this study was to develop a deep learning (DL) network for automated segmentation of TPM images, without significant loss in segmentation and myocardial velocity quantification accuracy compared with manual segmentation. We implemented a multi‐channel 3D (three dimensional; 2D + time) dense U‐Net that trained on magnitude and phase images and combined cross‐entropy, Dice, and Hausdorff distance loss terms to improve the segmentation accuracy and suppress unnatural boundaries. The dense U‐Net was trained and tested with 150 multi‐slice, multi‐phase TPM scans (114 scans for training, 36 for testing) from 99 heart transplant patients (44 females, 1‐4 scans/patient), where the magnitude and velocity‐encoded ( V x , V y , V z ) images were used as input and the corresponding manual segmentation masks were used as reference. The accuracy of DL segmentation was evaluated using quantitative metrics (Dice scores, Hausdorff distance) and linear regression and Bland‐Altman analyses on the resulting peak radial and longitudinal velocities ( V r and V z ). The mean segmentation time was about 2 h per patient for manual and 1.9 ± 0.3 s for DL. Our network produced good accuracy (median Dice = 0.85 for left ventricle (LV), 0.64 for right ventricle (RV), Hausdorff distance = 3.17 pixels) compared with manual segmentation. Peak V r and V z measured from manual and DL segmentations were strongly correlated ( R ≥ 0.88) and in good agreement with manual analysis (mean difference and limits of agreement for V z and V r were −0.05 ± 0.98 cm/s and −0.06 ± 1.18 cm/s for LV, and −0.21 ± 2.33 cm/s and 0.46 ± 4.00 cm/s for RV, respectively). The proposed multi‐channel 3D dense U‐Net was capable of reducing the segmentation time by 3,600‐fold, without significant loss in accuracy in tissue velocity measurements.
Hausdorff distance
Cite
Citations (5)
Segmentation of brain structures in a large dataset of magnetic resonance images (MRI) necessitates automatic segmentation instead of manual tracing. Automatic segmentation methods provide a much-needed alternative to manual segmentation which is both labor intensive and time-consuming. Among brain structures, the hippocampus presents a challenging segmentation task due to its irregular shape, small size, and unclear edges. In this work, we use T1-weighted MRI of 426 subjects to validate the approach and compare three automatic segmentation methods: FreeSurfer, LocalInfo, and ABSS. Four evaluation measures are used to assess agreement between automatic and manual segmentation of the hippocampus. ABSS outperformed the others based on the Dice coefficient, precision, Hausdorff distance, ASSD, RMS, similarity, sensitivity, and volume agreement. Moreover, comparison of the segmentation results, acquired using 1.5T and 3T MRI systems, showed that ABSS is more sensitive than the others to the field inhomogeneity of 3T MRI.
Hausdorff distance
Sørensen–Dice coefficient
Similarity (geometry)
Fully automatic
Tracing
Cite
Citations (1)
Chronic obstructive pulmonary disease (COPD) is closely related to the right ventricle and lung lobes. This study focuses on the segmentation of the right ventricle and lung lobes. We conducted experiments using the MMWHS and our lung lobe datasets and evaluated the segmentation using different training models. We observed that the multi-objective segmentation approach has advantages over single-objective segmentation in segmenting the right ventricle and lung lobes. For the segmentation of the right ventricle, the multi-objective segmentation approach yielded an improvement of 2.0% in the Dice coefficient and 2.5% in the Jaccard index compared to single-objective segmentation. For the segmentation of five lung lobes, the multi-objective segmentation outperformed the single-objective segmentation with Dice coefficient improvements of 1.4%, 1.0%, 1.5%, 0.7%, and 1.3%, respectively.
Sørensen–Dice coefficient
Jaccard index
Dice
Cite
Citations (1)
Segmentation of brain structures in a large dataset of magnetic resonance images (MRI) necessitates automatic segmentation instead of manual tracing. Automatic segmentation methods provide a much-needed alternative to manual segmentation which is both labor intensive and time-consuming. Among brain structures, the hippocampus presents a challenging segmentation task due to its irregular shape, small size, and unclear edges. In this work, we use T1-weighted MRI of 426 subjects to validate the approach and compare three automatic segmentation methods: FreeSurfer, LocalInfo, and ABSS. Four evaluation measures are used to assess agreement between automatic and manual segmentation of the hippocampus. ABSS outperformed the others based on the Dice coefficient, precision, Hausdorff distance, ASSD, RMS, similarity, sensitivity, and volume agreement. Moreover, comparison of the segmentation results, acquired using 1.5T and 3T MRI systems, showed that ABSS is more sensitive than the others to the field inhomogeneity of 3T MRI.
Hausdorff distance
Sørensen–Dice coefficient
Fully automatic
Similarity (geometry)
Tracing
Cite
Citations (0)
Segmentation of lungs with (large) lung cancer regions is a nontrivial problem. We present a new fully automated approach for segmentation of lungs with such high-density pathologies. Our method consists of two main processing steps. First, a novel robust active shape model (RASM) matching method is utilized to roughly segment the outline of the lungs. The initial position of the RASM is found by means of a rib cage detection method. Second, an optimal surface finding approach is utilized to further adapt the initial segmentation result to the lung. Left and right lungs are segmented individually. An evaluation on 30 data sets with 40 abnormal (lung cancer) and 20 normal left/right lungs resulted in an average Dice coefficient of 0.975±0.006 and a mean absolute surface distance error of 0.84±0.23 mm, respectively. Experiments on the same 30 data sets showed that our methods delivered statistically significant better segmentation results, compared to two commercially available lung segmentation approaches. In addition, our RASM approach is generally applicable and suitable for large shape models.
Sørensen–Dice coefficient
Position (finance)
Cite
Citations (183)
Precise segmentation of lung parenchyma is essential for effective analysis of the lung. Due to the obvious contrast and large regional area compared to other tissues in the chest, lung tissue is less difficult to segment. Special attention to details of lung segmentation is also needed. To improve the quality and speed of segmentation of lung parenchyma based on computed tomography (CT) or computed tomography angiography (CTA) images, the 4th International Symposium on Image Computing and Digital Medicine (ISICDM 2020) provides interesting and valuable research ideas and approaches. For the work of lung parenchyma segmentation, 9 of the 12 participating teams used the U-Net network or its modified forms, and others used the methods to improve the segmentation accuracy include attention mechanism, multi-scale feature information fusion. Among them, U-Net achieves the best results including that the final dice coefficient of CT segmentation is 0.991 and the final dice coefficient of CTA segmentation is 0.984. In addition, attention U-Net and nnU-Net network also performs well. In this paper, the methods chosen by 12 teams from different research groups are evaluated and their segmentation results are analyzed for the study and references to those involved.
Sørensen–Dice coefficient
Parenchyma
Dice
Feature (linguistics)
Cite
Citations (30)
Cite
Citations (0)
Short-axis MRI segmentation of the right ventricle plays an important role in assessing the structure and function of the right ventricle. However, RV segmentation is a challenge due to its complex crescent shape. In this paper, we propose a deep learning-based method for segmenting RV using the registration of the right ventricular shape model. The RV shape probability model is constructed using training samples. Next, aU-Net is trained using the shape prior probability by employing the registration technique. The shape model is registered to the network's predictive results to estimate a shape probability map, and a loss is defined as the Kullback-Leibler divergence between the prediction result and the shape probability map and the Kullback-Leibler divergence between the predictive result and the Ground-truth. The experimental results obtained from the cardiac automatic diagnosis challenge-medical imaging calculation and computer-aided intervention (ACDC-MICCAI) 2017 dataset show that the average 3D dice coefficient is 0.919, and the average 3D Hausdorff distance is 10.71mm. Our network has also been verified in the MICCAI2012 right ventricle segmentation challenge(RVSC) dataset. The average dice coefficient is 0.865, and the Hausdorff distance is 6. 10mm. The evaluation results show that our network outperforms the state-of-art methods in several evaluation indicators.
Hausdorff distance
Sørensen–Dice coefficient
Ground truth
Dice
Divergence (linguistics)
Cross entropy
Cite
Citations (1)
Effective image segmentation of cerebral structures is fundamental to 3-dimensional techniques such as augmented reality. To be clinically viable, segmentation algorithms should be fully automatic and easily integrated in existing digital infrastructure. We created a fully automatic adaptive-meshing-based segmentation system for T1-weighted magnetic resonance images (MRI) to automatically segment the complete ventricular system, running in a cloud-based environment that can be accessed on an augmented reality device. This study aims to assess the accuracy and segmentation time of the system by comparing it to a manually segmented ground truth dataset.A ground truth (GT) dataset of 46 contrast-enhanced and non-contrast-enhanced T1-weighted MRI scans was manually segmented. These scans also were uploaded to our system to create a machine-segmented (MS) dataset. The GT data were compared with the MS data using the Sørensen-Dice similarity coefficient and 95% Hausdorff distance to determine segmentation accuracy. Furthermore, segmentation times for all GT and MS segmentations were measured.Automatic segmentation was successful for 45 (98%) of 46 cases. Mean Sørensen-Dice similarity coefficient score was 0.83 (standard deviation [SD] = 0.08) and mean 95% Hausdorff distance was 19.06 mm (SD = 11.20). Segmentation time was significantly longer for the GT group (mean = 14405 seconds, SD = 7089) when compared with the MS group (mean = 1275 seconds, SD = 714) with a mean difference of 13,130 seconds (95% confidence interval 10,130-16,130).The described adaptive meshing-based segmentation algorithm provides accurate and time-efficient automatic segmentation of the ventricular system from T1 MRI scans and direct visualization of the rendered surface models in augmented reality.
Hausdorff distance
Sørensen–Dice coefficient
Ground truth
Similarity (geometry)
Mean-shift
Cite
Citations (11)
Purpose: Deformable image registration between CT and MRI remains a challenge due to variability in contrast, homogeneity, and field of view. This study aims to determine whether an optimized Mutual Information (MI) or Gradient Magnitude (GM) metric provides favorable results in the head and neck region. Methods: Head and neck CT and MR images were acquired for 14 patients as part of an MRI simulation study with a dual echo mDixon (in phase) T1 fast field echo sequence. Each pair of scans is registered with MRI as the reference image using the GM and MI metrics, where MI was optimized by controlling histogram parameters. Results are observed visually, and quantitatively measured by comparing Dice coefficient and Hausdorff distance of manually contoured bone structures. Results: Using the optimized GM metric, the average Dice coefficient is 0.52 ± 0.16; the average Hausdorff distance is 7.5 ± 1.9 mm. Using the optimized MI metric, the average Dice coefficient is 0.55 ± 0.085; the average Hausdorff distance is 6.6 ± 1.4 mm. Conclusion: While both metrics performed similarly, the MI metric has a higher Dice coefficient than the GM metric, and a lower Hausdorff distance. This suggests image registration with the Mutual Information metric may be preferred.
Hausdorff distance
Sørensen–Dice coefficient
Image registration
Dice
Similarity (geometry)
Cite
Citations (0)