Ground Truth Segmentation Research Articles

Automated Detection, Segmentation, and Tracking of Brain Metastases in Repeated Courses of Stereotactic Radiosurgery Using Integrated Artificial Intelligence.

Salvage stereotactic radiosurgery (SRS) for distant brain metastases has been demonstrated as a safe and effective approach for intracranial recurrences after initial SRS. However, accurate tumor detection and segmentation among responding tumors within the irradiated parenchyma can be challenging. The requirement for the registration and reference to the previous course of SRS is very time-consuming and suffers significant inter and intra-reader variability. Artificial intelligence (AI)-assisted system has been proven to improve the accuracy and efficiency in the clinical flow of de-novo SRS. We hypothesize that an integrated AI system can facilitate an automated tumor contouring process for repeated SRS. Three patients who underwent their third course of SRS to brain metastases were selected for the pioneering works. They have had two sessions of SRS with a mean lesion number of 4 and 3.7, respectively. VBrain, an FDA-approved brain tumor management AI platform, was used to co-registered serial MR scans and automatically identify, track, and contour brain metastases for each course of SRS. The AI also indicated new lesions and treated lesions for each course. Three radiation oncologists experienced in brain SRS contoured the gross tumor volumes (GTVs) of the third course of SRS in two reader modes (assisted then unassisted) with a memory washout period of one week between each section. The segmentation ground truth was established through consensus among the three experts. Lesion-wise sensitivity, contouring accuracy, and consuming time were compared between the two contouring modes. In each patient, there were 15, 11, and 9 metastases, with a median diameter of 4.72 (95% CI: 4.05, 6.91) mm. The mean lesion-wise sensitivity was 96.96±2.47% with AI assistance and 76.90 ± 7.10% without assistance. There were two false-positive lesions in the assisted read, resulting in a low average false-positive rate of 0.67 per patient, while no false positive for the unassisted mode. AI assistance improved contouring accuracy. The median Dice similarity coefficient (DSC) was 0.71 (95% CI: 0.55, 0.87) for assisted contouring and 0.65 (95% CI: 0.46, 0.85) for unassisted contouring. We also use average Hausdorff distance (HD) to measure segmentation results. The mean HD was 0.72± 0.13 mm versus 0.73±0.08 mm for the two contouring modes (p = 0.02) Furthermore, the median contouring time per case was significantly shorter with AI assistance than without assistance (20.8 minutes vs. 29.8 minutes; p < 0.001), corresponding to a 43.2% time-saving. Our results suggest that the integration of an AI-based system into repeated brain SRS can significantly improve the accuracy and efficiency of tumor detection and segmentation. This approach has the potential to streamline the treatment planning process for salvage SRS.

A Hybrid Intelligence (HI) System for Segmenting Rectoprostatic Spacer Gel and Key OARs on CT Images for Prostate Cancer Radiation Therapy Planning

Our hybrid intelligence (HI) system, combining natural and artificial intelligence, is effective for auto-contouring H&N and thorax organs at risk (OARs) for radiation therapy (RT) planning with FDA 510(k) clearance. The purpose of this study is to test the HI system to segment a commercially available retroprostatic hyaluronic acid spacer gel (RSG) and pelvic OARs in planning CT images for prostate cancer RT. HI can achieve clinically acceptable auto segmentation for tissue-equivalent RSG in this domain. RSG is injected in the peri-rectal space in men with prostate cancer prior to RT to minimize rectal toxicity. 190 patients with prostate cancer were included in this post-hoc image analysis from a multi-center, prospective, randomized trial, with 136 in the spacer arm. The HI system has 3 steps: rough recognition from fuzzy model (FM) based automatic anatomy recognition (AAR-R), deep learning-based recognition (DL-R) refinement, and deep learning-based delineation (DL-D) to contour objects guided by the recognition results. FM encodes high level 3D anatomy knowledge of object shape and its relationship with other OARs; DL-R and DL-D focus on pixel-level details. The 190 studies are divided into disjoint training (100) and testing (90) subsets. 100 samples are used in DL-R and DL-D training, with 45 to build the FM for AAR-R. RSG and 4 other OARs (pelvic skin, prostate, bladder, rectum) are contoured. Location error (LE) is used to evaluate recognition; Dice coefficient (DC) and Hausdorff distance (HD) are employed to evaluate delineation. Acceptability scores (AS) (range 1-5, 1 for poor quality, 5 for best quality) from an observer study are recorded for HI-output and ground truth masks of RSG for assessing segmentation quality. The HI system achieves highest DC (0.94±0.07) and lowest HD (1.96±1.61 mm) for bladder, for rectum and prostate similar DC (0.82±0.08) and HD (2.62±1.65mm), for RSG, the most challenging object, a good DC close to 0.7 (0.67±0.10) and excellent HD (2.66±1.44mm). AS for auto-segmentations (3.86±0.85) were significantly better than those for ground truth segmentations (3.45±1.00) (p = 0.02, paired t-test). Table 1 summarizes results. The HI system achieves clinically acceptable segmentations for pelvic OARs and significantly better acceptability of segmentation of RSG compared to clinically performed ground truth segmentations. This has implications in improving efficiency and accuracy of CT-based RT planning in patients with prostate cancer.

Automated measurements of interscrew angles in vertebral body tethering patients with deep learning

BACKGROUND CONTEXTVertebral body tethering is the most popular nonfusion treatment for adolescent idiopathic scoliosis. The effect of the tether cord on the spine can be segmentally assessed by comparing the angle between two adjacent screws (interscrew angle) over time. Tether breakage has historically been assessed radiographically by a change in adjacent interscrew angle by greater than 5° between two sets of imaging. A threshold for growth modulation has not yet been established in the literature. These angle measurements are time consuming and prone to interobserver variability. PURPOSEThe purpose of this study was to develop an automated deep learning algorithm for measuring the interscrew angle following VBT surgery. STUDY DESIGN/SETTINGSingle institution analysis of medical images. PATIENT SAMPLEWe analyzed 229 standing or bending AP or PA radiographs from 100 patients who had undergone VBT at our institution. OUTCOME MEASURESPhysiologic Measures: An image processing algorithm was used to measure interscrew angles. METHODSA total of 229 standing or bending AP or PA radiographs from 100 VBT patients with vertebral body tethers were identified. Vertebral body screws were segmented by hand for all images and interscrew angles measured manually for 60 of the included images. A U-Net deep learning model was developed to automatically segment the vertebral body screws. Screw label maps were used to develop and tune an image processing algorithm which measures interscrew angles. Finally, the completed model and algorithm pipeline was tested on a 30-image test set. Dice score and absolute error were used to measure performance. RESULTSInter- and Intra-rater reliability for manual angle measurements were assessed with ICC and were both 0.99. The segmentation model Dice score against manually segmented ground truth across the 30-image test set was 0.96. The average interscrew angle absolute error between the algorithm and manually measured ground truth was 0.66° and ranged from 0° to 2.67° in non-overlapping screws (N=206). The primary modes of failure for the model were overlapping screws on a right thoracic/left lumbar construct with two screws in one vertebra and overexposed images. An algorithm step which determines whether an overlapping screw was present correctly identified all overlapping screws, with no false positives. CONCLUSIONWe developed and validated an algorithm which measures interscrew angles for radiographs of vertebral body tether patients with an accuracy of within 1° for the majority of interscrew angles. The algorithm can process five images per second on a standard computer, leading to substantial time savings. This algorithm may be used for rapid processing of large radiographic databases of tether patients and could enable more rigorous definitions of growth modulation and cord breakage to be established.

Segmentation of Brain Metastases Using Background Layer Statistics (BLAST).

Accurate segmentation of brain metastases is important for treatment planning and evaluating response. The aim of this study was to assess the performance of a semiautomated algorithm for brain metastases segmentation using Background Layer Statistics (BLAST). Nineteen patients with 48 parenchymal and dural brain metastases were included. Segmentation was performed by 4 neuroradiologists and 1 radiation oncologist. K-means clustering was used to identify normal gray and white matter (background layer) in a 2D parameter space of signal intensities from postcontrast T2 FLAIR and T1 MPRAGE sequences. The background layer was subtracted and operator-defined thresholds were applied in parameter space to segment brain metastases. The remaining voxels were back-projected to visualize segmentations in image space and evaluated by the operators. Segmentation performance was measured by calculating the Dice-Sørensen coefficient and Hausdorff distance using ground truth segmentations made by the investigators. Contours derived from the segmentations were evaluated for clinical acceptance using a 5-point Likert scale. The median Dice-Sørensen coefficient was 0.82 for all brain metastases and 0.9 for brain metastases of ≥10 mm. The median Hausdorff distance was 1.4 mm. Excellent interreader agreement for brain metastases volumes was found with an intraclass correlation coefficient = 0.9978. The median segmentation time was 2.8 minutes/metastasis. Forty-five contours (94%) had a Likert score of 4 or 5, indicating that the contours were acceptable for treatment, requiring no changes or minor edits. We show accurate and reproducible segmentation of brain metastases using BLAST and demonstrate its potential as a tool for radiation planning and evaluating treatment response.

Anatomy-aware computed tomography-to-ultrasound spine registration.

Ultrasound (US) has demonstrated to be an effective guidance technique for lumbar spine injections, enabling precise needle placement without exposing the surgeon or the patient to ionizing radiation. However, noise and acoustic shadowing artifacts make US data interpretation challenging. To mitigate these problems, many authors suggested using computed tomography (CT)-to-US registration to align the spine in pre-operative CT to intra-operative US data, thus providing localization of spinallandmarks. In this paper, we propose a deep learning (DL) pipeline for CT-to-US registration and address the problem of a need for annotated medical data for network training. Firstly, we design a data generation method to generate paired CT-US data where the spine is deformed in a physically consistent manner. Secondly, we train a point cloud (PC) registration network using anatomy-aware losses to enforce anatomically consistentpredictions. Our proposed pipeline relies on training the network on realistic generated data. In our data generation method, we model the properties of the joints and disks between vertebrae based on biomechanical measurements in previous studies. We simulate the supine and prone position deformation by applying forces on the spine models. We choose the spine models from 35 patients in VerSe dataset. Each spine is deformed 10 times to create a noise-free data with ground-truth segmentation at hand. In our experiments, we use one-leave-out cross-validation strategy to measure the performance and the stability of the proposed method. For each experiment, we choose generated PCs from three spines as the test set. From the remaining, data from 3 spines act as the validation set and we use the rest of the data for training thealgorithm. To train our network, we introduce anatomy-aware losses and constraints on the movement to match the physics of the spine, namely, rigidity loss and bio-mechanical loss. We define rigidity loss based on the fact that each vertebra can only transform rigidly while the disks and the surrounding tissue are deformable. Second, by using bio-mechanical loss we stop the network from inferring extreme movements by penalizing the force needed to get to a certainpose. To validate the effectiveness of our fully automated data generation pipeline, we qualitatively assess the fidelity of the generated data. This assessment involves verifying the realism of the spinal deformation and subsequently confirming the plausibility of the simulated ultrasound images. Next, we demonstrate that the introduction of the anatomy-aware losses brings us closer to state-of-the-art (SOTA) and yields a reduction of 0.25 mm in terms of target registration error (TRE) compared to using only mean squared error (MSE) loss on the generated dataset. Furthermore, by using the proposed losses, the rigidity loss in inference decreases which shows that the inferred deformation respects the rigidity of the vertebrae and only introduces deformations in the soft tissue area to compensate the difference to the target PC. We also show that our results are close to the SOTA for the simulated US dataset with TRE of 3.89 mm and 3.63 mm for the proposed method and SOTA respectively. In addition, we show that our method is more robust against errors in the initialization in comparison to SOTA and significantly achieves better results (TRE of 4.88 mm compared to 5.66 mm) in this experiment. In conclusion, we present a pipeline for spine CT-to-US registration and explore the potential benefits of utilizing anatomy-aware losses to enhance registration results. Additionally, we propose a fully automatic method to synthesize paired CT-US data with physically consistent deformations, which offers the opportunity to generate extensive datasets for network training. The generated dataset and the source code for data generation and registration pipeline can be accessed via https://github.com/mfazampour/medphys_ct_us_registration.

LCW-Net: Low-light-image-based crop and weed segmentation network using attention module in two decoders

Crop segmentation using cameras is commonly used in large agricultural areas, but the time and duration of crop harvesting varies in large farms. Considering this situation, there is a need for low-light image-based segmentation of crop and weed images for late-time harvesting, but no prior research has considered this. As a first study on this topic, we propose a low-light image-based crop and weed segmentation network (LCW-Net) that uses an attention module in two decoders to perform only one step without restoration of low-light images. We also design a loss function to accurately segment regions of objects, crops, and weeds in low-light images to avoid training overfitting and balance the learning task for object, crop, and weed segmentation.There are no existing low-light public databases, and it is difficult to obtain ground truth segmentation information for self-collected database in low-light environments. Therefore, we experimented with converting two public databases, the crop and weed field image dataset (CWFID) and BoniRob dataset, into low-light datasets. The experimental results showed that the mean intersection of unions (mIoUs) of segmentation for crops and weeds were 0.8718 and 0.8693 for the BoniRob dataset, respectively, and 0.8337 and 0.8221 for the CWFID dataset, respectively, indicating that LCW-Net outperforms the state-of-the-art methods.

Automatic left ventricle segmentation via edge‐shape feature‐based fully convolutional neural network

AbstractLeft ventricle (LV) segmentation is essential to identify the cardiac functions for treating cardiovascular disorders. Cardiovascular magnetic resonance (CMRI) imaging is a non‐invasive technique for diagnosing cardiovascular diseases. CMRI is widely used to assess the functional integrity of the left and right ventricles for detecting changes in myocardial structure. Clinical parameters of the LV are often retrieved from CMRI scans, such as LV volumes, and ejection fraction. Moreover, manually segmenting cardiac diseases and evaluating such functions is time‐consuming and difficult for medical professionals. Deep learning networks require a lot of time, cost, and knowledge. To overcome this issue, a novel Edge and Shape feature‐based Fully Convolutional Neural Network (ES‐FCN) has been proposed for automatic LV segmentation. The ES‐FCN model segments MRI images based on edge maps instead of using gray‐scale images, which accelerates the performance of the FCN. The fuzzy‐based canny edge detection algorithm leverages fuzzy logic to detect structural changes in the LV and generate binary images. Additionally, binary‐valued kernels are used for convolution operations, where the binary values are influenced by biases derived from edge map shape descriptors. In ES‐FCN, bias values are learned from ground truth segmentation. The proposed ES‐FCN model achieves the Jaccard indices of 0.9484 ± 0.0188 for the ACDC dataset and 0.9476 ± 0.0237 for the LVSC dataset, and the dice index of 0.9319 ± 0.0188 for the ACDC dataset and 0.9314 ± 0.0237 for LVSC dataset respectively. The experimental results also reveal that the proposed ES‐FCN model is faster and requires minimal resources compared to state‐of‐the‐art models.

Deep Learning-Based Automated Magnetic Resonance Image Segmentation of the Lumbar Structure and Its Adjacent Structures at the L4/5 Level.

(1) Background: This study aims to develop a deep learning model based on a 3D Deeplab V3+ network to automatically segment multiple structures from magnetic resonance (MR) images at the L4/5 level. (2) Methods: After data preprocessing, the modified 3D Deeplab V3+ network of the deep learning model was used for the automatic segmentation of multiple structures from MR images at the L4/5 level. We performed five-fold cross-validation to evaluate the performance of the deep learning model. Subsequently, the Dice Similarity Coefficient (DSC), precision, and recall were also used to assess the deep learning model's performance. Pearson's correlation coefficient analysis and the Wilcoxon signed-rank test were employed to compare the morphometric measurements of 3D reconstruction models generated by manual and automatic segmentation. (3) Results: The deep learning model obtained an overall average DSC of 0.886, an average precision of 0.899, and an average recall of 0.881 on the test sets. Furthermore, all morphometry-related measurements of 3D reconstruction models revealed no significant difference between ground truth and automatic segmentation. Strong linear relationships and correlations were also obtained in the morphometry-related measurements of 3D reconstruction models between ground truth and automated segmentation. (4) Conclusions: We found it feasible to perform automated segmentation of multiple structures from MR images, which would facilitate lumbar surgical evaluation by establishing 3D reconstruction models at the L4/5 level.

Accurate mandibular canal segmentation of dental CBCT using a two-stage 3D-UNet based segmentation framework

ObjectivesThe objective of this study is to develop a deep learning (DL) model for fast and accurate mandibular canal (MC) segmentation on cone beam computed tomography (CBCT).MethodsA total of 220 CBCT scans from dentate subjects needing oral surgery were used in this study. The segmentation ground truth is annotated and reviewed by two senior dentists. All patients were randomly splitted into a training dataset (n = 132), a validation dataset (n = 44) and a test dataset (n = 44). We proposed a two-stage 3D-UNet based segmentation framework for automated MC segmentation on CBCT. The Dice Similarity Coefficient (DSC) and 95% Hausdorff Distance (95% HD) were used as the evaluation metrics for the segmentation model.ResultsThe two-stage 3D-UNet model successfully segmented the MC on CBCT images. In the test dataset, the mean DSC was 0.875 ± 0.045 and the mean 95% HD was 0.442 ± 0.379.ConclusionsThis automatic DL method might aid in the detection of MC and assist dental practitioners to set up treatment plans for oral surgery evolved MC.

Few-Shot Pixel-Precise Document Layout Segmentation via Dynamic Instance Generation and Local Thresholding.

Over the years, the humanities community has increasingly requested the creation of artificial intelligence frameworks to help the study of cultural heritage. Document Layout segmentation, which aims at identifying the different structural components of a document page, is a particularly interesting task connected to this trend, specifically when it comes to handwritten texts. While there are many effective approaches to this problem, they all rely on large amounts of data for the training of the underlying models, which is rarely possible in a real-world scenario, as the process of producing the ground truth segmentation task with the required precision to the pixel level is a very time-consuming task and often requires a certain degree of domain knowledge regarding the documents at hand. For this reason, in this paper, we propose an effective few-shot learning framework for document layout segmentation relying on two novel components, namely a dynamic instance generation and a segmentation refinement module. This approach is able of achieving performances comparable to the current state of the art on the popular Diva-HisDB dataset, while relying on just a fraction of the available data.

A systematic review of automated segmentation of 3D computed-tomography scans for volumetric body composition analysis.

Automated computed tomography (CT) scan segmentation (labelling of pixels according to tissue type) is now possible. This technique is being adapted to achieve three-dimensional (3D) segmentation of CT scans, opposed to single L3-slice alone. This systematic review evaluates feasibility and accuracy of automated segmentation of 3D CT scans for volumetric body composition (BC) analysis, as well as current limitations and pitfalls clinicians and researchers should be aware of. OVID Medline, Embase and grey literature databases up to October 2021 were searched. Original studies investigating automated skeletal muscle, visceral and subcutaneous AT segmentation from CT were included. Seven of the 92 studies met inclusion criteria. Variation existed in expertise and numbers of humans performing ground-truth segmentations used to train algorithms. There was heterogeneity in patient characteristics, pathology and CT phases that segmentation algorithms were developed upon. Reporting of anatomical CT coverage varied, with confusing terminology. Six studies covered volumetric regional slabs rather than the whole body. One study stated the use of whole-body CT, but it was not clear whether this truly meant head-to-fingertip-to-toe. Two studies used conventional computer algorithms. The latter five used deep learning (DL), an artificial intelligence technique where algorithms are similarly organized to brain neuronal pathways. Six of seven reported excellent segmentation performance (Dice similarity coefficients>0.9 per tissue). Internal testing on unseen scans was performed for only four of seven algorithms, whilst only three were tested externally. Trained DL algorithms achieved full CT segmentation in 12 to 75s versus 25min for non-DL techniques. DL enables opportunistic, rapid and automated volumetric BC analysis of CT performed for clinical indications. However, most CT scans do not cover head-to-fingertip-to-toe; further research must validate using common CT regions to estimate true whole-body BC, with direct comparison to single lumbar slice. Due to successes of DL, we expect progressive numbers of algorithms to materialize in addition to the seven discussed in this paper. Researchers and clinicians in the field of BC must therefore be aware of pitfalls. High Dice similarity coefficients do not inform the degree to which BC tissues may be under- or overestimated and nor does it inform on algorithm precision. Consensus is needed to define accuracy and precision standards for ground-truth labelling. Creation of a large international, multicentre common CT dataset with BC ground-truth labels from multiple experts could be a robust solution.

Graph-based automatic detection and classification of lesion changes in pairs of CT studies for oncology follow-up.

Radiological follow-up of oncology patients requires the quantitative analysis of lesion changes in longitudinal imaging studies, which is time-consuming, requires expertise, and is subject to variability. This paper presents a comprehensive graph-based method for the automatic detection and classification of lesion changes in current and prior CT scans. The inputs are the current and prior CT scans and their organ and lesion segmentations. Classification of lesion changes is formalized as bipartite graph matching where lesion pairings are computed by adaptive overlap-based lesion matching. Six types of lesion changes are computed by connected components analysis. The method was evaluated on 208 pairs of lung and liver CT scans from 57 patients with 4600 lesions, 1713 lesion matchings and 2887 lesion changes. Ground-truth lesion segmentations, lesion matchings and lesion changes were created by an expert radiologist. Our method yields a lesion matching rate accuracy of 99.7% (394/395) and 95.0% (1252/1318) for the lung and liver datasets. Precision and recall are > 0.99 and 0.94 and 0.95 (respectively) for the detection of lesion changes. The analysis of lesion changes helped the radiologist detect 48 missed lesions and 8 spurious lesions in the input ground-truth lesion datasets. The classification of lesion classification provides the clinician with a readily accessible and intuitive identification and classification of the lesion changes and their patterns in support of clinical decision making. Comprehensive automatic computer-aided lesion matching and analysis of lesion changes may improve quantitative follow-up and evaluation of disease status, assessment of treatment efficacy and response to therapy.

MixNet: Mix different networks for learning 3D implicit representations

We introduce a neural network, MixNet, for learning implicit representations of 3D subtle models with large smooth areas and exact shape details in the form of interpolation of two different implicit functions. Our network takes a point cloud as input and uses conventional MLP networks and SIREN networks to predict different implicit fields. We use a learnable interpolation function to combine the implicit values of these two networks and achieve the respective advantages of them. The network is self-supervised with only reconstruction loss, leading to faithful 3D reconstructions with smooth planes, correct details, and plausible spatial partition without any ground-truth segmentation. We evaluate our method on ABC, the largest and most diverse CAD dataset, and some typical shapes to test in terms of geometric correctness and surface smoothness to demonstrate superiority over current alternatives suitable for shape reconstruction.

Automated Radiomic Analysis of Vestibular Schwannomas and Inner Ears Using Contrast-Enhanced T1-Weighted and T2-Weighted Magnetic Resonance Imaging Sequences and Artificial Intelligence.

To objectively evaluate vestibular schwannomas (VSs) and their spatial relationships with the ipsilateral inner ear (IE) in magnetic resonance imaging (MRI) using deep learning. Cross-sectional study. A total of 490 adults with VS, high-resolution MRI scans, and no previous neurotologic surgery. MRI studies of VS patients were split into training (390 patients) and test (100 patients) sets. A three-dimensional convolutional neural network model was trained to segment VS and IE structures using contrast-enhanced T1-weighted and T2-weighted sequences, respectively. Manual segmentations were used as ground truths. Model performance was evaluated on the test set and on an external set of 100 VS patients from a public data set (Vestibular-Schwannoma-SEG). Dice score, relative volume error, average symmetric surface distance, 95th-percentile Hausdorff distance, and centroid locations. Dice scores for VS and IE volume segmentations were 0.91 and 0.90, respectively. On the public data set, the model segmented VS tumors with a Dice score of 0.89 ± 0.06 (mean ± standard deviation), relative volume error of 9.8 ± 9.6%, average symmetric surface distance of 0.31 ± 0.22 mm, and 95th-percentile Hausdorff distance of 1.26 ± 0.76 mm. Predicted VS segmentations overlapped with ground truth segmentations in all test subjects. Mean errors of predicted VS volume, VS centroid location, and IE centroid location were 0.05 cm 3 , 0.52 mm, and 0.85 mm, respectively. A deep learning system can segment VS and IE structures in high-resolution MRI scans with excellent accuracy. This technology offers promise to improve the clinical workflow for assessing VS radiomics and enhance the management of VS patients.

Toward Automated Detection of Silent Cerebral Infarcts in Children and Young Adults With Sickle Cell Anemia.

Silent cerebral infarcts (SCI) in sickle cell anemia (SCA) are associated with future strokes and cognitive impairment, warranting early diagnosis and treatment. Detection of SCI, however, is limited by their small size, especially when neuroradiologists are unavailable. We hypothesized that deep learning may permit automated SCI detection in children and young adults with SCA as a tool to identify the presence and extent of SCI in clinical and research settings. We utilized UNet-a deep learning model-for fully automated SCI segmentation. We trained and optimized UNet using brain magnetic resonance imaging from the SIT trial (Silent Infarct Transfusion). Neuroradiologists provided the ground truth for SCI diagnosis, while a vascular neurologist manually delineated SCI on fluid-attenuated inversion recovery and provided the ground truth for SCI segmentation. UNet was optimized for the highest spatial overlap between automatic and manual delineation (dice similarity coefficient). The optimized UNet was externally validated using an independent single-center prospective cohort of SCA participants. Model performance was evaluated through sensitivity and accuracy (%correct cases) for SCI diagnosis, dice similarity coefficient, intraclass correlation coefficient (metric of volumetric agreement), and Spearman correlation. The SIT trial (n=926; 31% with SCI; median age, 8.9 years) and external validation (n=80; 50% with SCI; age, 11.5 years) cohorts had small median lesion volumes of 0.40 and 0.25 mL, respectively. Compared with the neuroradiology diagnosis, UNet predicted SCI presence with 100% sensitivity and 74% accuracy. In magnetic resonance imaging with SCI, UNet reached a moderate spatial agreement (dice similarity coefficient, 0.48) and high volumetric agreement (intraclass correlation coefficient, 0.76; ρ=0.72; P<0.001) between automatic and manual segmentations. UNet, trained using a large pediatric SCA magnetic resonance imaging data set, sensitively detected small SCI in children and young adults with SCA. While additional training is needed, UNet may be integrated into the clinical workflow as a screening tool, aiding in SCI diagnosis.

Deep learning-aided extraction of outer aortic surface from CT angiography scans of patients with Stanford type B aortic dissection

BackgroundGuidelines recommend that aortic dimension measurements in aortic dissection should include the aortic wall. This study aimed to evaluate two-dimensional (2D)- and three-dimensional (3D)-based deep learning approaches for extraction of outer aortic surface in computed tomography angiography (CTA) scans of Stanford type B aortic dissection (TBAD) patients and assess the speed of different whole aorta (WA) segmentation approaches.MethodsA total of 240 patients diagnosed with TBAD between January 2007 and December 2019 were retrospectively reviewed for this study; 206 CTA scans from 206 patients with acute, subacute, or chronic TBAD acquired with various scanners in multiple different hospital units were included. Ground truth (GT) WAs for 80 scans were segmented by a radiologist using an open-source software. The remaining 126 GT WAs were generated via semi-automatic segmentation process in which an ensemble of 3D convolutional neural networks (CNNs) aided the radiologist. Using 136 scans for training, 30 for validation, and 40 for testing, 2D and 3D CNNs were trained to automatically segment WA. Main evaluation metrics for outer surface extraction and segmentation accuracy were normalized surface Dice (NSD) and Dice coefficient score (DCS), respectively.Results2D CNN outperformed 3D CNN in NSD score (0.92 versus 0.90, p = 0.009), and both CNNs had equal DCS (0.96 versus 0.96, p = 0.110). Manual and semi-automatic segmentation times of one CTA scan were approximately 1 and 0.5 h, respectively.ConclusionsBoth CNNs segmented WA with high DCS, but based on NSD, better accuracy may be required before clinical application. CNN-based semi-automatic segmentation methods can expedite the generation of GTs.Relevance statementDeep learning can speeds up the creation of ground truth segmentations. CNNs can extract the outer aortic surface in patients with type B aortic dissection.Key points• 2D and 3D convolutional neural networks (CNNs) can extract the outer aortic surface accurately.• Equal Dice coefficient score (0.96) was reached with 2D and 3D CNNs.• Deep learning can expedite the creation of ground truth segmentations.Graphical

Carotid atherosclerotic plaque segmentation in multi-weighted MRI using a two-stage neural network: advantages of training with high-resolution imaging and histology.

A reliable and automated method to segment and classify carotid artery atherosclerotic plaque components is needed to efficiently analyze multi-weighted magnetic resonance (MR) images to allow their integration into patient risk assessment for ischemic stroke. Certain plaque components such as lipid-rich necrotic core (LRNC) with hemorrhage suggest a greater likelihood of plaque rupture and stroke event. Assessment for presence and extent of LRNC could assist in directing treatment with impact upon patient outcomes. To address the need to accurately determine the presence and extent of plaque components on carotid plaque MRI, we proposed a two-staged deep-learning-based approach that consists of a convolutional neural network (CNN), followed by a Bayesian neural network (BNN). The rationale for the two-stage network approach is to account for the class imbalance of vessel wall and background by providing an attention mask to the BNN. A unique feature of the network training was to use ground truth defined by both high-resolution ex vivo MRI data and histopathology. More specifically, standard resolution 1.5 T in vivo MR image sets with corresponding high resolution 3.0 T ex vivo MR image sets and histopathology image sets were used to define ground-truth segmentations. Of these, data from 7 patients was used for training and from the remaining two was used for testing the proposed method. Next, to evaluate the generalizability of the method, we tested the method with an additional standard resolution 3.0 T in vivo data set of 23 patients obtained from a different scanner. Our results show that the proposed method yielded accurate segmentation of carotid atherosclerotic plaque and outperforms not only manual segmentation by trained readers, who did not have access to the ex vivo or histopathology data, but also three state-of-the-art deep-learning-based segmentation methods. Further, the proposed approach outperformed a strategy where the ground truth was generated without access to the high resolution ex vivo MRI and histopathology. The accurate performance of this method was also observed in the additional 23-patient dataset from a different scanner. In conclusion, the proposed method provides a mechanism to perform accurate segmentation of the carotid atherosclerotic plaque in multi-weighted MRI. Further, our study shows the advantages of using high-resolution imaging and histology to define ground truth for training deep-learning-based segmentation methods.

Accurate Bayesian segmentation of thalamic nuclei using diffusion MRI and an improved histological atlas

The human thalamus is a highly connected brain structure, which is key for the control of numerous functions and is involved in several neurological disorders. Recently, neuroimaging studies have increasingly focused on the volume and connectivity of the specific nuclei comprising this structure, rather than looking at the thalamus as a whole. However, accurate identification of cytoarchitectonically designed histological nuclei on standard in vivo structural MRI is hampered by the lack of image contrast that can be used to distinguish nuclei from each other and from surrounding white matter tracts. While diffusion MRI may offer such contrast, it has lower resolution and lacks some boundaries visible in structural imaging. In this work, we present a Bayesian segmentation algorithm for the thalamus. This algorithm combines prior information from a probabilistic atlas with likelihood models for both structural and diffusion MRI, allowing segmentation of 25 thalamic labels per hemisphere informed by both modalities. We present an improved probabilistic atlas, incorporating thalamic nuclei identified from histology and 45 white matter tracts surrounding the thalamus identified in ultra-high gradient strength diffusion imaging. We present a family of likelihood models for diffusion tensor imaging, ensuring compatibility with the vast majority of neuroimaging datasets that include diffusion MRI data. The use of these diffusion likelihood models greatly improves identification of nuclear groups versus segmentation based solely on structural MRI. Dice comparison of 5 manually identifiable groups of nuclei to ground truth segmentations show improvements of up to 10 percentage points. Additionally, our chosen model shows a high degree of reliability, with median test-retest Dice scores above 0.85 for four out of five nuclei groups, whilst also offering improved detection of differential thalamic involvement in Alzheimer’s disease (AUROC 81.98%). The probabilistic atlas and segmentation tool will be made publicly available as part of the neuroimaging package FreeSurfer (https://freesurfer.net/fswiki/ThalamicNucleiDTI).

Fully-automated detection of small bowel carcinoid tumors in CT scans using deep learning.

Small bowel carcinoid tumor is a rare neoplasm and increasing in incidence. Patients with small bowel carcinoid tumors often experience long delays in diagnosis due to the vague symptoms, slow growth of tumors, and lack of clinician awareness. Computed tomography (CT) is the most common imaging study for diagnosis of small bowel carcinoid tumor. It is often used with positron emission tomography (PET) to capture anatomical and functional aspects of carcinoid tumors and thus to increase thesensitivity. We compared three different kinds of methods for the automatic detection of small bowel carcinoid tumors on CT scans, which is the first to the best of ourknowledge. Thirty-three preoperative CT scans of 33 unique patients with surgically-proven carcinoid tumors within the small bowel were collected. Ground-truth segmentation of tumors was drawn on CT scans by referring to available 18 F-DOPA PET scans and the corresponding radiology report. These scans were split into the trainval set (n = 24) and the test positive set (n= 9). Additionally, 22 CT scans of 22 unique patients who had no evidence of the tumor were collected to comprise the test negative set. We compared three different kinds of detection methods, which are detection network, patch-based classification, and segmentation-based methods. We also investigated the usefulness of small bowel segmentation for reduction of false positives (FPs) for each method. Free-response receiver operating characteristic (FROC) curves and receiver operating characteristic (ROC) curves were used for lesion- and patient-level evaluations, respectively. Statistical analyses comparing the FROC and ROC curves were alsoperformed. The detection network method performed the best among the compared methods. For lesion-level detection, the detection network method, without the small bowel segmentation-based filtering, achieved sensitivity values of (60.8%, 81.1%, 82.4%, 86.5%) at per-scan FP rates of (1, 2, 4 ,8), respectively. The use of the small bowel segmentation did not improve the performance ( ). For patient-level detection, again the detection network method, but with the small bowel segmentation-based filtering, achieved the highest AUC of 0.86 with a sensitivity of 78% and specificity of 82% at the Youdenpoint. The carcinoid tumors in this patient population were very small and potentially difficult to diagnose. The presented method showed reasonable sensitivity at small numbers of FPs for lesion-level detection. It also achieved a promising AUC for patient-level detection. The method may have clinical application in patients with this rare and difficult to detectdisease.

Radiologic versus Segmentation Measurements to Quantify Wilms Tumor Volume on MRI in Pediatric Patients.

Wilms tumor is a common pediatric solid tumor. To evaluate tumor response to chemotherapy and decide whether nephron-sparing surgery is possible, tumor volume measurements based on magnetic resonance imaging (MRI) are important. Currently, radiological volume measurements are based on measuring tumor dimensions in three directions. Manual segmentation-based volume measurements might be more accurate, but this process is time-consuming and user-dependent. The aim of this study was to investigate whether manual segmentation-based volume measurements are more accurate and to explore whether these segmentations can be automated using deep learning. We included the MRI images of 45 Wilms tumor patients (age 0-18 years). First, we compared radiological tumor volumes with manual segmentation-based tumor volume measurements. Next, we created an automated segmentation method by training a nnU-Net in a five-fold cross-validation. Segmentation quality was validated by comparing the automated segmentation with the manually created ground truth segmentations, using Dice scores and the 95th percentile of the Hausdorff distances (HD95). On average, manual tumor segmentations result in larger tumor volumes. For automated segmentation, the median dice was 0.90. The median HD95 was 7.2 mm. We showed that radiological volume measurements underestimated tumor volume by about 10% when compared to manual segmentation-based volume measurements. Deep learning can potentially be used to replace manual segmentation to benefit from accurate volume measurements without time and observer constraints.