Three-dimensional U-Net Research Articles

CryoTEN: Efficiently Enhancing Cryo-EM Density Maps Using Transformers.

Cryogenic Electron Microscopy (cryo-EM) is a core experimental technique used to determine the structure of macromolecules such as proteins. However, the effectiveness of cryo-EM is often hindered by the noise and missing density values in cryo-EM density maps caused by experimental conditions such as low contrast and conformational heterogeneity. Although various global and local map sharpening techniques are widely employed to improve cryo-EM density maps, it is still challenging to efficiently improve their quality for building better protein structures from them. In this study, we introduce CryoTEN - a three-dimensional U-Net style transformer to improve cryo-EM maps effectively. CryoTEN is trained using a diverse set of 1,295 cryo-EM maps as inputs and their corresponding simulated maps generated from known protein structures as targets. An independent test set containing 150 maps is used to evaluate CryoTEN, and the results demonstrate that it can robustly enhance the quality of cryo-EM density maps. In addition, the automatic de novo protein structure modeling shows that the protein structures built from the density maps processed by CryoTEN have substantially better quality than those built from the original maps. Compared to the existing state-of-the-art deep learning methods for enhancing cryo-EM density maps, CryoTEN ranks second in improving the quality of density maps, while running > 10 times faster and requiring much less GPU memory than them. The source code and data is freely available at https://github.com/jianlin-cheng/cryoten.

Automatic segmentation of knee CT images of tibial plateau fractures based on three-dimensional U-Net: Assisting junior physicians with Schatzker classification

PurposeThis study aimed to automatically segment knee computed tomography (CT) images of tibial plateau fractures using a three-dimensional (3D) U-net-based method, accurately construct 3D maps of tibial plateau fractures, and examine their usefulness for Schatzker classification in clinical practice. MethodsWe retrospectively enrolled 234 cases with tibial plateau fractures from our hospital in this study. The four constituent bones of the knee were manually annotated using ITK-SNAP software. Finally, image features were extracted using deep learning. The usefulness of the results for Schatzker classification was examined by an orthopaedic and a radiology resident. ResultsOn average, our model required < 40 s to process a 3D CT scan of the knee. The average Dice coefficient for all four knee bones was higher than 0.950, and highly accurate 3D maps of the tibia were produced. With the aid of the results of our model, the accuracy, sensitivity, and specificity of the Schatzker classification of both residents improved. ConclusionsThe proposed method can rapidly and accurately segment knee CT images of tibial plateau fractures and assist residents with Schatzker classification, which can help improve diagnostic efficiency and reduce the workload of junior doctors in clinical practice.

Active Learning in Brain Tumor Segmentation with Uncertainty Sampling and Annotation Redundancy Restriction.

Deep learning models have demonstrated great potential in medical imaging but are limited by the expensive, large volume of annotations required. To address this, we compared different active learning strategies by training models on subsets of the most informative images using real-world clinical datasets for brain tumor segmentation and proposing a framework that minimizes the data needed while maintaining performance. Then, 638 multi-institutional brain tumor magnetic resonance imaging scans were used to train three-dimensional U-net models and compare active learning strategies. Uncertainty estimation techniques including Bayesian estimation with dropout, bootstrapping, and margins sampling were compared to random query. Strategies to avoid annotating similar images were also considered. We determined the minimum data necessary to achieve performance equivalent to the model trained on the full dataset (α = 0.05). Bayesian approximation with dropout at training and testing showed results equivalent to that of the full data model (target) with around 30% of the training data needed by random query to achieve target performance (p = 0.018). Annotation redundancy restriction techniques can reduce the training data needed by random query to achieve target performance by 20%. We investigated various active learning strategies to minimize the annotation burden for three-dimensional brain tumor segmentation. Dropout uncertainty estimation achieved target performance with the least annotated data.

Deep Learning to Differentiate Benign and Malignant Vertebral Fractures at Multidetector CT.

Background Differentiating between benign and malignant vertebral fractures poses diagnostic challenges. Purpose To investigate the reliability of CT-based deep learning models to differentiate between benign and malignant vertebral fractures. Materials and Methods CT scans acquired in patients with benign or malignant vertebral fractures from June 2005 to December 2022 at two university hospitals were retrospectively identified based on a composite reference standard that included histopathologic and radiologic information. An internal test set was randomly selected, and an external test set was obtained from an additional hospital. Models used a three-dimensional U-Net encoder-classifier architecture and applied data augmentation during training. Performance was evaluated using the area under the receiver operating characteristic curve (AUC) and compared with that of two residents and one fellowship-trained radiologist using the DeLong test. Results The training set included 381 patients (mean age, 69.9 years ± 11.4 [SD]; 193 male) with 1307 vertebrae (378 benign fractures, 447 malignant fractures, 482 malignant lesions). Internal and external test sets included 86 (mean age, 66.9 years ± 12; 45 male) and 65 (mean age, 68.8 years ± 12.5; 39 female) patients, respectively. The better-performing model of two training approaches achieved AUCs of 0.85 (95% CI: 0.77, 0.92) in the internal and 0.75 (95% CI: 0.64, 0.85) in the external test sets. Including an uncertainty category further improved performance to AUCs of 0.91 (95% CI: 0.83, 0.97) in the internal test set and 0.76 (95% CI: 0.64, 0.88) in the external test set. The AUC values of residents were lower than that of the best-performing model in the internal test set (AUC, 0.69 [95% CI: 0.59, 0.78] and 0.71 [95% CI: 0.61, 0.80]) and external test set (AUC, 0.70 [95% CI: 0.58, 0.80] and 0.71 [95% CI: 0.60, 0.82]), with significant differences only for the internal test set (P < .001). The AUCs of the fellowship-trained radiologist were similar to those of the best-performing model (internal test set, 0.86 [95% CI: 0.78, 0.93; P = .39]; external test set, 0.71 [95% CI: 0.60, 0.82; P = .46]). Conclusion Developed models showed a high discriminatory power to differentiate between benign and malignant vertebral fractures, surpassing or matching the performance of radiology residents and matching that of a fellowship-trained radiologist. © RSNA, 2024 See also the editorial by Booz and D'Angelo in this issue.

Using an artificial intelligence model to detect and localize visible clinically significant prostate cancer in prostate magnetic resonance imaging: a multicenter external validation study.

An increasing number of patients with suspected clinically significant prostate cancer (csPCa) are undergoing prostate multiparametric magnetic resonance imaging (mpMRI). The role of artificial intelligence (AI) algorithms in interpreting prostate mpMRI needs to be tested with multicenter external data. This study aimed to investigate the diagnostic efficacy of an AI model in detecting and localizing visible csPCa on mpMRI a multicenter external data set. The data of 2,105 patients suspected of having prostate cancer from four hospitals were retrospectively collected to develop an AI model to detect and localize suspicious csPCa. The lesions were annotated based on pathology records by two radiologists. Diffusion-weighted imaging (DWI) and apparent diffusion coefficient (ADC) values were used as the input for the three-dimensional U-Net framework. Subsequently, the model was validated using an external data set comprising the data of 557 patients from three hospitals. Sensitivity, specificity, and accuracy were employed to evaluate the diagnostic efficacy of the model. At the lesion level, the model had a sensitivity of 0.654. At the overall sextant level, the model had a sensitivity, specificity, and accuracy of 0.846, 0.884, and 0.874, respectively. At the patient level, the model had a sensitivity, specificity, and accuracy of 0.943, 0.776, and 0.849, respectively. The AI-predicted accuracy for the csPCa patients (231/245, 0.943) was significantly higher than that for the non-csPCa patients (242/312, 0.776) (P<0.001). The lesion number and tumor volume were greater in the correctly diagnosed patients than the incorrectly diagnosed patients (both P<0.001). Among the positive patients, those with lower average ADC values had a higher rate of correct diagnosis than those with higher average ADC values (P=0.01). The AI model exhibited acceptable accuracy in detecting and localizing visible csPCa at the patient and sextant levels. However, further improvements need to be made to enhance the sensitivity of the model at the lesion level.

Enhancing facial feature de-identification in multiframe brain images: A generative adversarial network approach.

The collection of head images for public datasets in the field of brain science has grown remarkably in recent years, underscoring the need for robust de-identification methods to adhere with privacy regulations. This paper elucidates a novel deep learning-based approach to deidentifying facial features in brain images using a generative adversarial network to synthesize new facial features and contours. We employed the precision of the three-dimensional U-Net model to detect specific features such as the ears, nose, mouth, and eyes. Results: Our method diverges from prior studies by highlighting partial regions of the head image rather than comprehensive full-head images. We trained and tested our model on a dataset comprising 490 cases from a publicly available head computed tomography image dataset and an additional 70 cases with head MR images. Integrated data proved advantageous, with promising results. The nose, mouth, and eye detection achieved 100% accuracy, while ear detection reached 85.03% in the training dataset. In the testing dataset, ear detection accuracy was 65.98%, and the validation dataset ear detection attained 100%. Analysis of pixel value histograms demonstrated varying degrees of similarity, as measured by the Structural Similarity Index (SSIM), between raw and generated features across different facial features. The proposed methodology, tailored for partial head image processing, is well suited for real-world imaging examination scenarios and holds potential for future clinical applications contributing to the advancement of research in de-identification technologies, thus fortifying privacy safeguards.

Deep-learning-driven dose prediction and verification for stereotactic radiosurgical treatment of isolated brain metastases

PurposeWhile deep learning has shown promise for automated radiotherapy planning, its application to the specific scenario of stereotactic radiosurgery (SRS) for brain metastases using fixed-field intensity modulated radiation therapy (IMRT) on a linear accelerator remains limited. This work aimed to develop and verify a deep learning-guided automated planning protocol tailored for this scenario.MethodsWe collected 70 SRS plans for solitary brain metastases, of which 36 cases were for training and 34 for testing. Test cases were derived from two distinct clinical institutions. The envisioned automated planning process comprised (1): clinical dose prediction facilitated by deep-learning algorithms (2); transformation of the forecasted dose into executable plans via voxel-centric dose emulation (3); validation of the envisaged plan employing a precise dosimeter in conjunction with a linear accelerator. Dose prediction paradigms were established by engineering and refining two three-dimensional UNet architectures (UNet and AttUNet). Input parameters encompassed computed tomography scans from clinical plans and demarcations of the focal point alongside organs at potential risk (OARs); the ensuing output manifested as a 3D dose matrix tailored for each case under scrutiny.ResultsDose estimations rendered by both models mirrored the manual plans and adhered to clinical stipulations. As projected by the dual models, the apex and average doses for OARs did not deviate appreciably from those delineated in the manual plan (P-value≥0.05). AttUNet showed promising results compared to the foundational UNet. Predicted doses showcased a pronounced dose gradient, with peak concentrations localized within the target vicinity. The executable plans conformed to clinical dosimetric benchmarks and aligned with their associated verification assessments (100% gamma approval rate at 3 mm/3%).ConclusionThis study demonstrates an automated planning technique for fixed-field IMRT-based SRS for brain metastases. The envisaged plans met clinical requirements, were reproducible across centers, and achievable in deliveries. This represents progress toward automated paradigms for this specific scenario.

Fully Automated Segmentation of Human Eyeball Using Three-Dimensional U-Net in T2 Magnetic Resonance Imaging.

To develop and validate a fully automated deep-learning-based tool for segmentation of the human eyeball using a three-dimensional (3D) U-Net, compare its performance to semiautomatic segmentation ground truth and a two-dimensional (2D) U-Net, and analyze age and sex differences in eyeball volume, as well as gaze-dependent volume consistency in normal subjects. We retrospectively collected 474 magnetic resonance imaging (MRI) scans, including different gazing scans, from 119 patients. A 10-fold cross-validation was applied to separate the dataset into training, test, and validation sets for both the 3D U-Net and 2D U-Net. Performance accuracy was measured using four quantitative metrics compared to the ground truth, and Bland-Altman plot analysis was conducted. Age and sex differences in eyeball volume and variability in eyeball volume differences across gazing directions were analyzed. The 3D U-Net outperformed the 2D U-Net with mean accuracy scores >0.95, showing acceptable agreement in the Bland-Altman plot analysis despite a tendency for slight overestimation (mean difference = -0.172 cm³). Significant sex differences and age effects on eyeball volume were observed for both methods (P < 0.05). No significant volume differences were found between the segmentation methods or within each method for the different gazing directions. Significant differences in performance accuracy were identified among the five gazing directions, with the upward direction showing a notably lower performance. Our study demonstrated the effectiveness of 3D U-Net human eyeball volume segmentation using T2-weighted MRI. The robustness and reliability of 3D U-Net across diverse populations and gaze directions support enhanced ophthalmic diagnosis and treatment strategies. Our findings demonstrate the feasibility of using the proposed 3D U-Net model for the automatic segmentation of the human eyeball, with potential applications in various ophthalmic research fields that require the analysis of 3D geometric eye globe shapes or eye movement detection.

A deep learning-based automated algorithm for labeling coronary arteries in computed tomography angiography images

ObjectiveUsing two three-dimensional U-Net architectures for myocardium structure extraction and a distance transformation algorithm specifically for the left circumflex artery, we have designed a fully automated algorithm for coronary artery labeling in coronary computed tomography angiography (CCTA) images.MethodsIn this retrospective analysis, a cohort of 157 patients who had undergone coronary computed tomography angiography (CCTA) was included. An automated coronary artery labeling algorithm was developed using a distance transformation approach to delineate the anatomical segments along the centerlines extracted from the CCTA images. A total of 16 segments were successfully identified and labeled. The algorithm’s outcomes were recorded and reviewed by three experts, and the performance of segment detection and labeling was assessed. Additionally, the level of agreement in manually labeled segments between two experts was quantified.ResultsWhen comparing the labels generated by the experts with those produced by the algorithm, it was necessary to modify or eliminate 117 labels (5.4%) out of 2180 segments assigned by the algorithm. The overall accuracy for label presence was 96.2%, with an average overlap of 94.0% between the expert reference and algorithm-generated labels. Furthermore, the average agreement rate between the two experts stood at 95.0%.ConclusionsBased on the labels of the clinical experts, the proposed deep learning algorithm exhibits high accuracy for automatic labeling. Therefore, our proposed method exhibits promising results for the automatic labeling of the coronary arteries and will alleviate the burden on radiologists in the near future.

Intelligent noninvasive meningioma grading with a fully automatic segmentation using interpretable multiparametric deep learning.

To establish a robust interpretable multiparametric deep learning (DL) model for automatic noninvasive grading of meningiomas along with segmentation. In total, 257 patients with pathologically confirmed meningiomas (162 low-grade, 95 high-grade) who underwent a preoperative brain MRI, including T2-weighted (T2) and contrast-enhanced T1-weighted images (T1C), were included in the institutional training set. A two-stage DL grading model was constructed for segmentation and classification based on multiparametric three-dimensional U-net and ResNet. The models were validated in the external validation set consisting of 61 patients with meningiomas (46 low-grade, 15 high-grade). Relevance-weighted Class Activation Mapping (RCAM) method was used to interpret the DL features contributing to the prediction of the DL grading model. On external validation, the combined T1C and T2 model showed a Dice coefficient of 0.910 in segmentation and the highest performance for meningioma grading compared to the T2 or T1C only models, with an area under the curve (AUC) of 0.770 (95% confidence interval: 0.644-0.895) and accuracy, sensitivity, and specificity of 72.1%, 73.3%, and 71.7%, respectively. The AUC and accuracy of the combined DL grading model were higher than those of the human readers (AUCs of 0.675-0.690 and accuracies of 65.6-68.9%, respectively). The RCAM of the DL grading model showed activated maps at the surface regions of meningiomas indicating that the model recognized the features at the tumor margin for grading. An interpretable multiparametric DL model combining T1C and T2 can enable fully automatic grading of meningiomas along with segmentation. • The multiparametric DL model showed robustness in grading and segmentation on external validation. • The diagnostic performance of the combined DL grading model was higher than that of the human readers. • The RCAM interpreted that DL grading model recognized the meaningful features at the tumor margin for grading.

Deep Learning for the Automatic Segmentation of Extracranial Venous Malformations of the Head and Neck from MR Images Using 3D U-Net.

Background: It is difficult to characterize extracranial venous malformations (VMs) of the head and neck region from magnetic resonance imaging (MRI) manually and one at a time. We attempted to perform the automatic segmentation of lesions from MRI of extracranial VMs using a convolutional neural network as a deep learning tool. Methods: T2-weighted MRI from 53 patients with extracranial VMs in the head and neck region was used for annotations. Preprocessing management was performed before training. Three-dimensional U-Net was used as a segmentation model. Dice similarity coefficients were evaluated along with other indicators. Results: Dice similarity coefficients in 3D U-Net were found to be 99.75% in the training set and 60.62% in the test set. The models showed overfitting, which can be resolved with a larger number of objects, i.e., MRI VM images. Conclusions: Our pilot study showed sufficient potential for the automatic segmentation of extracranial VMs through deep learning using MR images from VM patients. The overfitting phenomenon observed will be resolved with a larger number of MRI VM images.

DSA Image Analysis of Clinical Features and Nursing Care of Cerebral Aneurysm Patients Based on the Deep Learning Algorithm.

Objective A deep learning algorithm was developed for automatic detection and localization of intracranial aneurysms in DSA, and its clinical characteristics were analyzed, and targeted nursing measures were formulated. Methods Using a retrospective multicenter study method based on radiology reports, DSA images of aneurysms were randomly divided into 75 cases in the training set, 20 cases in the internal test set, and 35 cases in the external test set. Using a computer-aided detection method based on the three-dimensional U-Net (3D U-Net), after preprocessing DSA images, automatic segmentation of intracranial blood vessels is performed to obtain regions of interest, and based on the segmentation results, physicians' annotations are introduced. The 3D U-Net network model is trained and adjusted, and the obtained model is used to automatically detect the cerebral aneurysm area. Results Fivefold cross-validation was used for the training set and the internal test set, and a sensitivity of (94.4 ± 1.1)% was obtained. Automatic detection of aneurysms was performed on the external test set, and the average false positive rate was 0.86 FPs/case (false positives/case). The resulting sensitivity was 82.9%. The classification comparison of external test sets showed that the sensitivity of the method for detecting aneurysms with sizes of 5.00~<10.00 mm and ≥10.00 mm (88.2% and 100.0%) was higher than that for aneurysms with sizes of <3.00 mm and 3.00~<5.00 mm (50.0% and 72.7%). The sensitivity of patients aged 50-60 years and >60 years (90.0% and 87.5%) was higher than that of patients aged <50 years (66.7%), and there was little difference between different genders (84.6% in males and 81.8% in females). Conclusion The deep learning algorithm has high diagnostic performance in detecting intracranial aneurysms, which is verified by external datasets.

Comparison of two-dimensional and three-dimensional U-Net architectures for segmentation of adipose tissue in cardiac magnetic resonance images.

The process of identifying cardiac adipose tissue (CAT) from volumetric magnetic resonance imaging of the heart is tedious, time-consuming, and often dependent on observer interpretation. Many 2-dimensional (2D) convolutional neural networks (CNNs) have been implemented to automate the cardiac segmentation process, but none have attempted to identify CAT. Furthermore, the results from automatic segmentation of other cardiac structures leave room for improvement. This study investigated the viability of a 3-dimensional (3D) CNN in comparison to a similar 2D CNN. Both models used a U-Net architecture to simultaneously classify CAT, left myocardium, left ventricle, and right myocardium. The multi-phase model trained with multiple observers' segmentations reached a whole-volume Dice similarity coefficient (DSC) of 0.925 across all classes and 0.640 for CAT specifically; the corresponding 2D model's DSC across all classes was 0.902 and 0.590 for CAT specifically. This 3D model also achieved a higher level of CAT-specific DSC agreement with a group of observers with a Williams Index score of 0.973 in comparison to the 2D model's score of 0.822.

Efficient multiscale fully convolutional UNet model for segmentation of 3D lung nodule from CT image.

Purpose: Segmentation of lung nodules in chest CT images is essential for image-driven lung cancer diagnosis and follow-up treatment planning. Manual segmentation of lung nodules is subjective because the approach depends on the knowledge and experience of the specialist. We proposed a multiscale fully convolutional three-dimensional UNet (MF-3D UNet) model for automatic segmentation of lung nodules in CT images. Approach: The proposed model employs two strategies, fusion of multiscale features with Maxout aggregation and trainable downsampling, to improve the performance of nodule segmentation in 3D CT images. The fusion of multiscale (fine and coarse) features with the Maxout function allows the model to retain the most important features while suppressing the low-contribution features. The trainable downsampling process is used instead of fixed pooling-based downsampling. Results: The performance of the proposed MF-3D UNet model is examined by evaluating the model with CT scans obtained from the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) dataset. A quantitative and visual comparative analysis of the proposed work with various customized UNet models is also presented. The comparative analysis shows that the proposed model yields reliable segmentation results compared with other methods. The experimental result of 3D MF-UNet shows encouraging results in the segmentation of different types of nodules, including juxta-pleural, solitary pulmonary, and non-solid nodules, with an average Dice similarity coefficient of , and it outperforms other CNN-based segmentation models. Conclusions: The proposed model accurately segments the nodules using multiscale feature aggregation and trainable downsampling approaches. Also, 3D operations enable precise segmentation of complex nodules using inter-slice connections.

Deep Learning Automated Background Phase Error Correction for Abdominopelvic 4D Flow MRI.

Background Four-dimensional (4D) flow MRI has the potential to provide hemodynamic insights for a variety of abdominopelvic vascular diseases, but its clinical utility is currently impaired by background phase error, which can be challenging to correct. Purpose To assess the feasibility of using deep learning to automatically perform image-based background phase error correction in 4D flow MRI and to compare its effectiveness relative to manual image-based correction. Materials and Methods A convenience sample of 139 abdominopelvic 4D flow MRI acquisitions performed between January 2016 and July 2020 was retrospectively collected. Manual phase error correction was performed using dedicated imaging software and served as the reference standard. After reserving 40 examinations for testing, the remaining examinations were randomly divided into training (86% [85 of 99]) and validation (14% [14 of 99]) data sets to train a multichannel three-dimensional U-Net convolutional neural network. Flow measurements were obtained for the infrarenal aorta, common iliac arteries, common iliac veins, and inferior vena cava. Statistical analyses included Pearson correlation, Bland-Altman analysis, and F tests with Bonferroni correction. Results A total of 139 patients (mean age, 47 years ± 14 [standard deviation]; 108 women) were included. Inflow-outflow correlation improved after manual correction (ρ = 0.94, P < .001) compared with that before correction (ρ = 0.50, P < .001). Automated correction showed similar results (ρ = 0.91, P < .001) and demonstrated very strong correlation with manual correction (ρ = 0.98, P < .001). Both correction methods reduced inflow-outflow variance, improving mean difference from -0.14 L/min (95% limits of agreement: -1.61, 1.32) (uncorrected) to 0.05 L/min (95% limits of agreement: -0.32, 0.42) (manually corrected) and 0.05 L/min (95% limits of agreement: -0.38, 0.49) (automatically corrected). There was no significant difference in inflow-outflow variance between manual and automated correction methods (P = .10). Conclusion Deep learning automated phase error correction reduced inflow-outflow bias and variance of volumetric flow measurements in four-dimensional flow MRI, achieving results comparable with manual image-based phase error correction. © RSNA, 2021 See also the editorial by Roldán-Alzate and Grist in this issue.

Interinstitutional Portability of a Deep Learning Brain MRI Lesion Segmentation Algorithm.

To assess how well a brain MRI lesion segmentation algorithm trained at one institution performed at another institution, and to assess the effect of multi-institutional training datasets for mitigating performance loss. In this retrospective study, a three-dimensional U-Net for brain MRI abnormality segmentation was trained on data from 293 patients from one institution (IN1) (median age, 54 years; 165 women; patients treated between 2008 and 2018) and tested on data from 51 patients from a second institution (IN2) (median age, 46 years; 27 women; patients treated between 2003 and 2019). The model was then trained on additional data from various sources: (a) 285 multi-institution brain tumor segmentations, (b) 198 IN2 brain tumor segmentations, and (c) 34 IN2 lesion segmentations from various brain pathologic conditions. All trained models were tested on IN1 and external IN2 test datasets, assessing segmentation performance using Dice coefficients. The U-Net accurately segmented brain MRI lesions across various pathologic conditions. Performance was lower when tested at an external institution (median Dice score, 0.70 [IN2] vs 0.76 [IN1]). Addition of 483 training cases of a single pathologic condition, including from IN2, did not raise performance (median Dice score, 0.72; P = .10). Addition of IN2 training data with heterogeneous pathologic features, representing only 10% (34 of 329) of total training data, increased performance to baseline (Dice score, 0.77; P < .001). This final model produced total lesion volumes with a high correlation to the reference standard (Spearman r = 0.98). For brain MRI lesion segmentation, adding a modest amount of relevant training data from an external institution to a previously trained model supported successful application of the model to this external institution.Keywords: Neural Networks, Brain/Brain Stem, Segmentation Supplemental material is available for this article. © RSNA, 2021.

Three-dimensional U-Net Convolutional Neural Network for Detection and Segmentation of Intracranial Metastases.

To develop and validate a neural network for automated detection and segmentation of intracranial metastases on brain MRI studies obtained for stereotactic radiosurgery treatment planning. In this retrospective study, 413 patients (average age, 61 years ± 12 [standard deviation]; 238 women) with a total of 5202 intracranial metastases (median volume, 0.05 cm3; interquartile range, 0.02-0.18 cm3) undergoing stereotactic radiosurgery at one institution were included (January 2017 to February 2020). A total of 563 MRI examinations were performed among the patients, and studies were split into training (n = 413), validation (n = 50), and test (n = 100) datasets. A three-dimensional (3D) U-Net convolutional network was trained and validated on 413 T1 postcontrast or subtraction scans, and several loss functions were evaluated. After model validation, 100 discrete test patients, who underwent imaging after the training and validation patients, were used for final model evaluation. Performance for detection and segmentation of metastases was evaluated using Dice scores, false discovery rates, and false-negative rates, and a comparison with neuroradiologist interrater reliability was performed. The median Dice score for segmenting enhancing metastases in the test set was 0.75 (interquartile range, 0.63-0.84). There were strong correlations between manually segmented and predicted metastasis volumes (r = 0.98, P < .001) and between the number of manually segmented and predicted metastases (R = 0.95, P < .001). Higher Dice scores were strongly correlated with larger metastasis volumes on a logarithmically transformed scale (r = 0.71). Sensitivity across the whole test sample was 70.0% overall and 96.4% for metastases larger than 6 mm. There was an average of 0.46 false-positive results per scan, with the positive predictive value being 91.5%. In comparison, the median Dice score between two neuroradiologists was 0.85 (interquartile range, 0.80-0.89), with sensitivity across the test sample being 87.9% overall and 98.4% for metastases larger than 6 mm. A 3D U-Net-based convolutional neural network was able to segment brain metastases with high accuracy and perform detection at the level of human interrater reliability for metastases larger than 6 mm.Keywords: Adults, Brain/Brain Stem, CNS, Feature detection, MR-Imaging, Neural Networks, Neuro-Oncology, Quantification, Segmentation© RSNA, 2021.

Automated Segmentation of Infarct Lesions in T1-Weighted MRI Scans Using Variational Mode Decomposition and Deep Learning.

Automated segmentation methods are critical for early detection, prompt actions, and immediate treatments in reducing disability and death risks of brain infarction. This paper aims to develop a fully automated method to segment the infarct lesions from T1-weighted brain scans. As a key novelty, the proposed method combines variational mode decomposition and deep learning-based segmentation to take advantages of both methods and provide better results. There are three main technical contributions in this paper. First, variational mode decomposition is applied as a pre-processing to discriminate the infarct lesions from unwanted non-infarct tissues. Second, overlapped patches strategy is proposed to reduce the workload of the deep-learning-based segmentation task. Finally, a three-dimensional U-Net model is developed to perform patch-wise segmentation of infarct lesions. A total of 239 brain scans from a public dataset is utilized to develop and evaluate the proposed method. Empirical results reveal that the proposed automated segmentation can provide promising performances with an average dice similarity coefficient (DSC) of 0.6684, intersection over union (IoU) of 0.5022, and average symmetric surface distance (ASSD) of 0.3932, respectively.

An algorithm for three-dimensional plumonary parenchymal segmentation by integrating surfacelet transform with pulse coupled neural network

In order to overcome the difficulty in lung parenchymal segmentation due to the factors such as lung disease and bronchial interference, a segmentation algorithm for three-dimensional lung parenchymal is presented based on the integration of surfacelet transform and pulse coupled neural network (PCNN). First, the three-dimensional computed tomography of lungs is decomposed into surfacelet transform domain to obtain multi-scale and multi-directional sub-band information. The edge features are then enhanced by filtering sub-band coefficients using local modified Laplacian operator. Second, surfacelet inverse transform is implemented and the reconstructed image is fed back to the input of PCNN. Finally, iteration process of the PCNN is carried out to obtain final segmentation result. The proposed algorithm is validated on the samples of public dataset. The experimental results demonstrate that the proposed algorithm has superior performance over that of the three-dimensional surfacelet transform edge detection algorithm, the three-dimensional region growing algorithm, and the three-dimensional U-NET algorithm. It can effectively suppress the interference coming from lung lesions and bronchial, and obtain a complete structure of lung parenchyma.

Development and Validation of a Modified Three-Dimensional U-Net Deep-Learning Model for Automated Detection of Lung Nodules on Chest CT Images From the Lung Image Database Consortium and Japanese Datasets

A more accurate lung nodule detection algorithm is needed. We developed a modified three-dimensional (3D) U-net deep-learning model for the automated detection of lung nodules on chest CT images. The purpose of this study was to evaluate the accuracy of the developed modified 3D U-net deep-learning model. In this Health Insurance Portability and Accountability Act-compliant, Institutional Review Board-approved retrospective study, the 3D U-net based deep-learning model was trained using the Lung Image Database Consortium and Image Database Resource Initiative dataset. For internal model validation, we used 89 chest CT scans that were not used for model training. For external model validation, we used 450 chest CT scans taken at an urban university hospital in Japan. Each case included at least one nodule of >5 mm identified by an experienced radiologist. We evaluated model accuracy using the competition performance metric (CPM) (average sensitivity at 1/8, 1/4, 1/2, 1, 2, 4, and 8 false-positives per scan). The 95% confidence interval (CI) was computed by bootstrapping 1000 times. In the internal validation, the CPM was 94.7% (95% CI: 89.1%-98.6%). In the external validation, the CPM was 83.3% (95% CI: 79.4%-86.1%). The modified 3D U-net deep-learning model showed high performance in both internal and external validation.