Higher Dice Coefficient Research Articles

Verdiff-Net: A Conditional Diffusion Framework for Spinal Medical Image Segmentation.

Spinal medical image segmentation is critical for diagnosing and treating spinal disorders. However, ambiguity in anatomical boundaries and interfering factors in medical images often cause segmentation errors. Current deep learning models cannot fully capture the intrinsic data properties, leading to unstable feature spaces. To tackle the above problems, we propose Verdiff-Net, a novel diffusion-based segmentation framework designed to improve segmentation accuracy and stability by learning the underlying data distribution. Verdiff-Net integrates a multi-scale fusion module (MSFM) for fine feature extraction and a noise semantic adapter (NSA) to refine segmentation masks. Validated across four multi-modality spinal datasets, Verdiff-Net achieves a high Dice coefficient of 93%, demonstrating its potential for clinical applications in precision spinal surgery.

Designing a TPMS metamaterial via deep learning and topology optimization

Data-driven models that act as surrogates for computationally costly 3D topology optimization techniques are very popular because they help alleviate multiple time-consuming 3D finite element analyses during optimization. In this study, one such 3D CNN-based surrogate model for the topology optimization of Schoen’s gyroid triply periodic minimal surface unit cell is investigated. Gyroid-like unit cells are designed using a voxel algorithm and homogenization-based topology optimization codes in MATLAB. A few such optimization data are used as input–output for supervised learning of the topology-optimization process via the 3D CNN model in Python code. These models could then be used to instantaneously predict the optimized unit cell geometry for any topology parameters. The high accuracy of the model was demonstrated by a low mean square error metric and a high Dice coefficient metric. The model has the major disadvantage of running numerous costly topology optimization runs but has the advantages that the trained model can be reused for different cases of TO and that the methodology of the accelerated design of 3D metamaterials can be extended for designing any complex, computationally costly problems of metamaterials with multi-objective properties or multiscale applications. The main purpose of this paper is to provide the complete associated MATLAB and PYTHON codes for optimizing the topology of any cellular structure and predicting new topologies using deep learning for educational purposes.

A Fully Automated Pipeline Using Swin Transformers for Deep Learning-Based Blood Segmentation on Head Computed Tomography Scans After Aneurysmal Subarachnoid Hemorrhage

Accurate volumetric assessment of spontaneous aneurysmal subarachnoid hemorrhage (aSAH) is a labor-intensive task performed with current manual and semiautomatic methods that might be relevant for its clinical and prognostic implications. In the present research, we sought to develop and validate an artificial intelligence-driven, fully automated blood segmentation tool for subarachnoid hemorrhage (SAH) patients via noncontrast computed tomography (NCCT) scans employing a transformer-based Swin-UNETR architecture. We retrospectively analyzed NCCT scans from patients with confirmed aSAH utilizing the Swin-UNETR for segmentation. The performance of the proposed method was evaluated against manually segmented ground truth data using metrics such as Dice score, intersection over union, volumetric similarity index , symmetric average surface distance , sensitivity, and specificity. A validation cohort from an external institution was included to test the generalizability of the model. The model demonstrated high accuracy with robust performance metrics across the internal and external validation cohorts. Notably, it achieved high Dice coefficient (0.873±0.097), intersection over union (0.810±0.092), volumetric similarity index (0.840±0.131), sensitivity (0.821±0.217), and specificity (0.996±0.004) values and a low symmetric average surface distance (1.866±2.910), suggesting proficiency in segmenting blood in SAH patients. The model's efficiency was reflected in its processing speed, indicating potential for real-time applications. Our Swin UNETR-based model offers significant advances in the automated segmentation of blood in SAH patients on NCCT images. Despite the computational demands, the model operates effectively on standard hardware with a user-friendly interface, facilitating broader clinical adoption. Further validation across diverse datasets is warranted to confirm its clinical reliability.

Evaluating normalized registration and preprocessing methodologies for the analysis of brain MRI in pediatric patients with shunt-treated hydrocephalus.

Registration to a standardized template (i.e. "normalization") is a critical step when performing neuroimaging studies. We present a comparative study involving the evaluation of general-purpose registration algorithms for pediatric patients with shunt treated hydrocephalus. Our sample dataset presents a number of intersecting challenges for registration, representing the potentially large deformations to both brain structures and overall brain shape, artifacts from shunts, and morphological differences corresponding to age. The current study assesses the normalization accuracy of shunt-treated hydrocephalus patients using freely available neuroimaging registration tools. Anatomical neuroimages from eight pediatric patients with shunt-treated hydrocephalus were normalized. Four non-linear registration algorithms were assessed in addition to the preprocessing steps of skull-stripping and bias-correction. Registration accuracy was assessed using the Dice Coefficient (DC) and Hausdorff Distance (HD) in subcortical and cortical regions. A total of 592 registrations were performed. On average, normalizations performed using the brain extracted and bias-corrected images had a higher DC and lower HD compared to full head/ non-biased corrected images. The most accurate registration was achieved using SyN by ANTs with skull-stripped and bias corrected images. Without preprocessing, the DARTEL Toolbox was able to produce normalized images with comparable accuracy. The use of a pediatric template as an intermediate registration did not improve normalization. Using structural neuroimages from patients with shunt-treated pediatric hydrocephalus, it was demonstrated that there are tools which perform well after specified pre-processing steps were taken. Overall, these results provide insight to the performance of registration programs that can be used for normalization of brains with complex pathologies.

A novel convolutional neural network for identification of retinal layers using sliced optical coherence tomography images

Retinal imaging is crucial for observing the retina and accurately diagnosing pathological problems. Optical Coherence Tomography (OCT) has been a transformative breakthrough for developing high-resolution cross-sectional images. It is imperative to delineate the multiple layers of the retina for a proper diagnosis. A novel segmentation-based approach is introduced in this study to identify seven distinct layers of the retina using OCT images. The developed approach presents SliceOCTNet, a customized U-shaped Convolutional Neural Network (CNN) that introduces group normalization and intricate skip connections. Paired alongside a hybrid loss function, the SliceOCTNet outperformed most state-of-the-art approaches. The introduction of Group Normalization in SliceOCTNet stabilized the model and improved layer identification even when working with small datasets. The use of skip connections also contributed to an improvement in the spatial outlook of the model. Implementing a hybrid loss function addresses the class imbalance problem in the dataset. Duke University’s spectral-domain optical coherence tomography (SD-OCT) B-scan dataset of healthy and Diabetic Macular Edema (DME) afflicted patients was utilized to train and evaluate the SliceOCTNet. The model accurately recognizes the seven layers of the retina. It can achieve a high dice coefficient value of 0.941 and refine the segmentation process to a higher level of precision.

Soybean root image dataset and its deep learning application for nodule segmentation

The root is very important for anchoring the plant and acquiring water and nutrients from the soil for growth. Plant root trait phenotyping using imaging (2D and 3D) techniques is gaining importance in agriculture to enhance the breeding of superior cultivars. However, most root image data have been collected under controlled growth conditions, such as in greenhouses and growth chambers. In this study, we propose collecting soybean root image datasets from open field conditions for root phenotyping. There are numerous methods for segmenting images, but many of them are not suitable for the tiny and nodulated architecture of root. Advancements such as deep learning (DL) methods and semantic segmentation algorithms have been actively applied in many research studies. In the current study, we used convolutional neural network (CNN) based U-Net with six convolution methods and Deeplabv3+ and compared their performance in nodule identification based on the Dice coefficient, GPU memory usage, qualitative image segmentation, training time, and inference time. Also, we employed two image processing, resizing and patching, for efficient training of images with very high resolution. The results indicated accurate segmentation of root nodules; in addition, the results for the segmentation of root nodules using DL high-resolution (HR) images were more precise and efficient than those with low-resolution (LR) images since the Dice coefficient value and rate of nodule size were high in HR images although the former required more training time and higher GPU memory. Furthermore, in comparison to the resizing approach, patch-based approach outperformed in all aspects of performance. Among the DL models that we employed for our study, the grouped convolution-based U-net showed the best results having high Dice coefficient values 0.647 and 0.618 for 600x400 and 300x200 images respectively in resizing approach and dynamic convolution-based U-net showed the best results having Dice coefficient values 0.792 and 0.777 for 600x400 and 300x200 patch images respectively in patch-based approach. Thus, the optimized DL model presented in this study has the potential to be a valuable tool for the automated analysis of soybean root systems, assisting in the measurement of various root quality parameters and contributing to the development of superior cultivars.

Evaluation of computed tomography images under deep learning in the diagnosis of severe pulmonary infection.

This work aimed to explore the diagnostic value of a deep convolutional neural network (CNN) combined with computed tomography (CT) images in patients with severe pneumonia complicated with pulmonary infection. A total of 120 patients with severe pneumonia complicated by pulmonary infection admitted to the hospital were selected as research subjects and underwent CT imaging scans. The empty convolution (EC) and U-net phase were combined to construct an EC-U-net, which was applied to process the CT images. The results showed that the learning rate of the EC-U-net model decreased substantially with increasing training times until it stabilized and reached zero after 40 training times. The segmentation result of the EC-U-net model for the CT image was very similar to that of the mask image, except for some deviations in edge segmentation. The EC-U-net model exhibited a significantly smaller cross-entropy loss function (CELF) and a higher Dice coefficient than the CNN algorithm. The diagnostic accuracy of CT images based on the EC-U-net model for severe pneumonia complicated with pulmonary infection was substantially higher than that of CT images alone, while the false negative rate (FNR) and false positive rate (FPR) were substantially lower (P < 0.05). Moreover, the true positive rates (TPRs) of CT images based on the EC-U-net model for patchy high-density shadows, diffuse ground glass density shadows, pleural effusion, and lung consolidation were obviously higher than those of the original CT images (P < 0.05). In short, the EC-U-net model was superior to the traditional algorithm regarding the overall performance of CT image segmentation, which can be clinically applied. CT images based on the EC-U-net model can clearly display pulmonary infection lesions, improve the clinical diagnosis of severe pneumonia complicated with pulmonary infection, and help to screen early pulmonary infection and carry out symptomatic treatment.

Evaluation of a fully automated anatomical left atrial mapping pipeline enabled by artficial intelligence

Abstract Funding Acknowledgements Type of funding sources: Public Institution(s). Main funding source(s): Dr. Fabbricatore and Dr Valeriano have been supported by a research grant provided by the Cardiopath PhD program (University of Naples Federico II) Background Pulmonary vein isolation (PVI) is increasingly performed globally for the treatment of atrial fibrillation [1]. Using a standardized and reproducible workflow, we previously reported an average procedure time of 76.1 ± 26.2 min [2]. Over the last few years, due to workflow optimization and new technological developments, we were able to reduce the skin-to-skin time towards 47.3 ± 7.1 min (unpublished data). Our workflow implements rotational angiography (RA) of the left atrium (LA), which is segmented manually, then converted to a 3D mesh and finally imported into the electroanatomical mapping system (EAM). A second 3D mesh is created by cutting the left bottom anterior part of the LA, including the left atrial appendage, in order to clearly visualize the ridge (Figure 1). Recently, we developed an integrated artificial intelligence (AI) pipeline to automate this process (Figure 2). OBJECTIVE The aim of the study was to evaluate the efficiency of the new AI-based fully automated anatomical mapping of the LA and the efficacy of the automated segmentation. Methods The last 50 PVI procedures that were manually segmented in our lab were processed once more using the AI pipeline. The efficiency of the AI pipeline was assessed by comparing the elapsed time, for both our current manual workflow and the AI pipeline, between the reconstruction of the computed tomography (CT) volume from the 2D RA and the moment both 3D meshes are imported into the EAM. The efficacy of the pipeline was analyzed by calculating the Dice coefficient between the manual and automated segmentation as well as by visually grading segmentation quality on a 4-point Likert scale (1: Bad,2: Sufficient, 3: Good, 4: Excellent) by an experienced cardiac electrophysiologist. Significance testing of timing and visual grading differences between both workflows was performed using a one-sided Wilcoxon signed-rank test. Variables were summarized using median (Q1-Q3). Results The time to segment the LA, create the 3D meshes and import them into the EAM was significantly reduced from 560.0 (513.0–636.5) seconds to 52.0 (49.0–56.0) seconds for the manual and automated workflow respectively (p&amp;lt;0.0001). This reduction is equivalent to about 18% of the current skin-to-skin time for PVI in our lab. Furthermore, we reported a high Dice coefficient between the automated and manual segmentations valued at 0.928 (0.917–0.939). Finally, the subjective quality of the automated segmentation was superior compared to the manual segmentation, with respective values of 4 (3–4) and 3 (3–3.5) (p=0.006). Conclusion We developed an efficient and effective AI pipeline enabling fully automated anatomical mapping of the LA, potentially reducing our current PVI skin-to-skin time by 18%. The robustness of the AI segmentation model has been validated objectively (Dice coefficient) and subjectively. The efficacy of the AI cut model, the second model in the pipeline, remains to be verified.

ARA-net: an attention-aware retinal atrophy segmentation network coping with fundus images.

Accurately detecting and segmenting areas of retinal atrophy are paramount for early medical intervention in pathological myopia (PM). However, segmenting retinal atrophic areas based on a two-dimensional (2D) fundus image poses several challenges, such as blurred boundaries, irregular shapes, and size variation. To overcome these challenges, we have proposed an attention-aware retinal atrophy segmentation network (ARA-Net) to segment retinal atrophy areas from the 2D fundus image. In particular, the ARA-Net adopts a similar strategy as UNet to perform the area segmentation. Skip self-attention connection (SSA) block, comprising a shortcut and a parallel polarized self-attention (PPSA) block, has been proposed to deal with the challenges of blurred boundaries and irregular shapes of the retinal atrophic region. Further, we have proposed a multi-scale feature flow (MSFF) to challenge the size variation. We have added the flow between the SSA connection blocks, allowing for capturing considerable semantic information to detect retinal atrophy in various area sizes. The proposed method has been validated on the Pathological Myopia (PALM) dataset. Experimental results demonstrate that our method yields a high dice coefficient (DICE) of 84.26%, Jaccard index (JAC) of 72.80%, and F1-score of 84.57%, which outperforms other methods significantly. Our results have demonstrated that ARA-Net is an effective and efficient approach for retinal atrophic area segmentation in PM.

INTEGRATING PROSTATE SPECIFIC ANTIGEN DENSITY BIOMARKER INTO DEEP LEARNING PROSTATE MRI LESION SEGMENTATION MODELS.

Prostate cancer lesion segmentation in multi-parametric magnetic resonance imaging (mpMRI) is crucial for pre-biopsy diagnosis and targeted biopsy guidance. Deep convolution neural networks have been widely utilized for lesion segmentation. However, these methods fail to achieve a high Dice coefficient because of the large variations in lesion size and location within the gland. To address this problem, we integrate the clinically-meaningful prostate specific antigen density (PSAD) biomarker into the deep learning model using feature-wise transformations to condition the features in latent space, and thus control the size of lesion prediction. We tested our models on a public dataset with 214 annotated mpMRI scans and compared the segmentation performance to a baseline 3D U-Net model. Results demonstrate that integrating the PSAD biomarker significantly improves segmentation performance in both Dice coefficient and centroid distance metric.

UMRFormer-net: a three-dimensional U-shaped pancreas segmentation method based on a double-layer bridged transformer network.

Methods based on the combination of transformer and convolutional neural networks (CNNs) have achieved impressive results in the field of medical image segmentation. However, most of the recently proposed combination segmentation approaches simply treat transformers as auxiliary modules which help to extract long-range information and encode global context into convolutional representations, and there is a lack of investigation on how to optimally combine self-attention with convolution. We designed a novel transformer block (MRFormer) that combines a multi-head self-attention layer and a residual depthwise convolutional block as the basic unit to deeply integrate both long-range and local spatial information. The MRFormer block was embedded between the encoder and decoder in U-Net at the last two layers. This framework (UMRFormer-Net) was applied to the segmentation of three-dimensional (3D) pancreas, and its ability to effectively capture the characteristic contextual information of the pancreas and surrounding tissues was investigated. Experimental results show that the proposed UMRFormer-Net achieved accuracy in pancreas segmentation that was comparable or superior to that of existing state-of-the-art 3D methods in both the Clinical Proteomic Tumor Analysis Consortium Pancreatic Ductal Adenocarcinoma (CPTAC-PDA) dataset and the public Medical Segmentation Decathlon dataset (self-division). UMRFormer-Net statistically significantly outperformed existing transformer-related methods and state-of-the-art 3D methods (P<0.05, P<0.01, or P<0.001), with a higher Dice coefficient (85.54% and 77.36%, respectively) or a lower 95% Hausdorff distance (4.05 and 8.34 mm, respectively). UMRFormer-Net can obtain more matched and accurate segmentation boundary and region information in pancreas segmentation, thus improving the accuracy of pancreas segmentation. The code is available at https://github.com/supersunshinefk/UMRFormer-Net.

A hybrid attentional guidance network for tumors segmentation of breast ultrasound images.

In recent years, breast cancer has become the greatest threat to women. There are many studies dedicated to the precise segmentation of breast tumors, which is indispensable in computer-aided diagnosis. Deep neural networks have achieved accurate segmentation of images. However, convolutional layers are biased to extract local features and tend to lose global and location information as the network deepens, which leads to a decrease in breast tumors segmentation accuracy. For this reason, we propose a hybrid attention-guided network (HAG-Net). We believe that this method will improve the detection rate and segmentation of tumors in breast ultrasound images. The method is equipped with multi-scale guidance block (MSG) for guiding the extraction of low-resolution location information. Short multi-head self-attention (S-MHSA) and convolutional block attention module are used to capture global features and long-range dependencies. Finally, the segmentation results are obtained by fusing multi-scale contextual information. We compare with 7 state-of-the-art methods on two publicly available datasets through five random fivefold cross-validations. The highest dice coefficient, Jaccard Index and detect rate ([Formula: see text]%, [Formula: see text]%, [Formula: see text]% and [Formula: see text]%, [Formula: see text]%, [Formula: see text]%, separately) obtained on two publicly available datasets(BUSI and OASUBD), prove the superiority of our method. HAG-Net can better utilize multi-resolution features to localize the breast tumors. Demonstrating excellent generalizability and applicability for breast tumors segmentation compare to other state-of-the-art methods.

TrEnD: A transformer-based encoder-decoder model with adaptive patch embedding for mass segmentation in mammograms.

Breast cancer is one of the most prevalent malignancies diagnosed in women. Mammogram inspection in the search and delineation of breast tumors is an essential prerequisite for a reliable diagnosis. However, analyzing mammograms by radiologists is time-consuming and prone to errors. Therefore, the development of computer-aided diagnostic (CAD) systems to automate the mass segmentation procedure is greatly expected. Accurate breast mass segmentation in mammograms remains challenging in CAD systems due to the low contrast, various shapes, and fuzzy boundaries of masses. In this paper, we propose a fully automatic and effective mass segmentation model based on deep learning for improving segmentation performance. We propose an effective transformer-based encoder-decoder model (TrEnD). Firstly, we introduce a lightweight method for adaptive patch embedding (APE) of the transformer, which utilizes superpixels to adaptively adjust the size and position of each patch. Secondly, we introduce a hierarchical transformer-encoder and attention-gated-decoder structure, which is beneficial for progressively suppressing interference feature activations in irrelevant background areas. Thirdly, a dual-branch design is employed to extract and fuse globally coarse and locally fine features in parallel, which could capture the global contextual information and ensure the relevance and integrity of local information. The model is evaluated on two public datasets CBIS-DDSM and INbreast. To further demonstrate the robustness of TrEnD, different cropping strategies are applied to these datasets, termed tight, loose, maximal, and mix-frame. Finally, ablation analysis is performed to assess the individual contribution of each module to the model performance. The proposed segmentation model provides a high Dice coefficient and Intersection over Union (IoU) of 92.20% and 85.81% on the mix-frame CBIS-DDSM, while 91.83% and 85.29% for the mix-frame INbreast, respectively. The segmentation performance outperforms the current state-of-the-art approaches. By adding the APE and attention-gated module, the Dice and IoU have improved by 6.54% and 10.07%. According to extensive qualitative and quantitative assessments, the proposed network is effective for automatic breast mass segmentation, and has adequate potential to offer technical assistance for subsequent clinical diagnoses.

Dense-PSP-UNet: A neural network for fast inference liver ultrasound segmentation

Liver Ultrasound (US) or sonography is popularly used because of its real-time output, low-cost, ease-of-use, portability, and non-invasive nature. Segmentation of real-time liver US is essential for diagnosing and analyzing liver conditions (e.g., hepatocellular carcinoma (HCC)), assisting the surgeons/radiologists in therapeutic procedures. In this paper, we propose a method using a modified Pyramid Scene Parsing (PSP) module in tuned neural network backbones to achieve real-time segmentation without compromising the segmentation accuracy. Considering widespread noise in US data and its impact on outcomes, we study the impact of pre-processing and the influence of loss functions on segmentation performance. We have tested our method after annotating a publicly available US dataset containing 2400 images of 8 healthy volunteers (link to the annotated dataset is provided); the results show that the Dense-PSP-UNet model achieves a high Dice coefficient of 0.913±0.024 while delivering a real-time performance of 37 frames per second (FPS).

Reference Data Accuracy Impacts Burned Area Product Validation: The Role of the Expert Analyst

Accurate reference data to validate burned area (BA) products are crucial to obtaining reliable accuracy metrics for such products. However, the accuracy of reference data can be affected by numerous factors; hence, we can expect some degree of deviation with respect to real ground conditions. Since reference data are usually produced by semi-automatic methods, where human-based image interpretation is an important part of the process, in this study, we analyze the impact of the interpreter on the accuracy of the reference data. Here, we compare the accuracy metrics of the FireCCI51 BA product obtained from reference datasets that were produced by different analysts over 60 sites located in tropical regions of South America. Additionally, fire severity, tree cover percentage, and canopy height were selected as explanatory sources of discrepancies between interpreters’ reference BA classifications. We found significant differences between the FireCCI51 accuracy metrics obtained with the different reference datasets. The highest accuracies (highest Dice coefficient) were obtained with the reference dataset produced by the most experienced interpreter. The results indicated that fire severity is the main source of discrepancy between interpreters. Disagreement between interpreters was more likely to occur in areas with low fire severity. We conclude that the training and experience of the interpreter play a crucial role in guaranteeing the quality of the reference data.

Automated segmentation and feature discovery of age-related macular degeneration and Stargardt disease via self-attended neural networks

Age-related macular degeneration (AMD) and Stargardt disease are the leading causes of blindness for the elderly and young adults respectively. Geographic atrophy (GA) of AMD and Stargardt atrophy are their end-stage outcomes. Efficient methods for segmentation and quantification of these atrophic lesions are critical for clinical research. In this study, we developed a deep convolutional neural network (CNN) with a trainable self-attended mechanism for accurate GA and Stargardt atrophy segmentation. Compared with traditional post-hoc attention mechanisms which can only visualize CNN features, our self-attended mechanism is embedded in a fully convolutional network and directly involved in training the CNN to actively attend key features for enhanced algorithm performance. We applied the self-attended CNN on the segmentation of AMD and Stargardt atrophic lesions on fundus autofluorescence (FAF) images. Compared with a preexisting regular fully convolutional network (the U-Net), our self-attended CNN achieved 10.6% higher Dice coefficient and 17% higher IoU (intersection over union) for AMD GA segmentation, and a 22% higher Dice coefficient and a 32% higher IoU for Stargardt atrophy segmentation. With longitudinal image data having over a longer time, the developed self-attended mechanism can also be applied on the visual discovery of early AMD and Stargardt features.

Phase-based masking for quantitative susceptibility mapping of the human brain at 9.4T.

To develop improved tissue masks for QSM. Masks including voxels at the brain surface were automatically generated from the magnitude alone (MM) or combined with test functions from the first (PG) or second (PB) derivative of the sign of the wrapped phase. Phase images at 3T and 9.4T were simulated at different TEs and used to generate a mask, PItoh , with between-voxel phase differences less than π. MM, PG, and PB were compared with PItoh . QSM were generated from 3D multi-echo gradient-echo data acquired at 9.4T (21 subjects aged: 20-56y), and from the QSM2016 challenge 3T data using different masks, unwrapping, background removal, and dipole inversion algorithms. QSM contrast was quantified using age-based iron concentrations. Close to air cavities, phase wraps became denser with increasing field and echo time, yielding increased values of the test functions. Compared with PItoh , PB had the highest Dice coefficient, while PG had the lowest and MM the highest percentage of voxels outside PItoh. Artifacts observed in QSM at 9.4T with MM were mitigated by stronger background filters but yielded a reduced QSM contrast. With PB, QSM contrast was greater and artifacts diminished. Similar results were obtained with challenge data, evidencing larger effects of mask close to air cavities. Automatic, phase-based masking founded on the second derivative of the sign of the wrapped phase, including cortical voxels at the brain surface, was able to mitigate artifacts and restore QSM contrast across cortical and subcortical brain regions.

Artificial intelligence for stepwise diagnosis and monitoring of COVID-19.

BackgroundMain challenges for COVID-19 include the lack of a rapid diagnostic test, a suitable tool to monitor and predict a patient’s clinical course and an efficient way for data sharing among multicenters. We thus developed a novel artificial intelligence system based on deep learning (DL) and federated learning (FL) for the diagnosis, monitoring, and prediction of a patient’s clinical course.MethodsCT imaging derived from 6 different multicenter cohorts were used for stepwise diagnostic algorithm to diagnose COVID-19, with or without clinical data. Patients with more than 3 consecutive CT images were trained for the monitoring algorithm. FL has been applied for decentralized refinement of independently built DL models.ResultsA total of 1,552,988 CT slices from 4804 patients were used. The model can diagnose COVID-19 based on CT alone with the AUC being 0.98 (95% CI 0.97–0.99), and outperforms the radiologist’s assessment. We have also successfully tested the incorporation of the DL diagnostic model with the FL framework. Its auto-segmentation analyses co-related well with those by radiologists and achieved a high Dice’s coefficient of 0.77. It can produce a predictive curve of a patient’s clinical course if serial CT assessments are available.InterpretationThe system has high consistency in diagnosing COVID-19 based on CT, with or without clinical data. Alternatively, it can be implemented on a FL platform, which would potentially encourage the data sharing in the future. It also can produce an objective predictive curve of a patient’s clinical course for visualization.Key Points• CoviDet could diagnose COVID-19 based on chest CT with high consistency; this outperformed the radiologist’s assessment. Its auto-segmentation analyses co-related well with those by radiologists and could potentially monitor and predict a patient’s clinical course if serial CT assessments are available. It can be integrated into the federated learning framework.• CoviDet can be used as an adjunct to aid clinicians with the CT diagnosis of COVID-19 and can potentially be used for disease monitoring; federated learning can potentially open opportunities for global collaboration.Supplementary InformationThe online version contains supplementary material available at 10.1007/s00330-021-08334-6.

Clinical validation of the use of prototype software for automatic cartilage segmentation to quantify knee cartilage in volunteers

BackgroundThe cartilage segmentation algorithms make it possible to accurately evaluate the morphology and degeneration of cartilage. There are some factors (location of cartilage subregions, hydrarthrosis and cartilage degeneration) that may influence the accuracy of segmentation. It is valuable to evaluate and compare the accuracy and clinical value of volume and mean T2* values generated directly from automatic knee cartilage segmentation with those from manually corrected results using prototype software.MethodThirty-two volunteers were recruited, all of whom underwent right knee magnetic resonance imaging examinations. Morphological images were obtained using a three-dimensional (3D) high-resolution Double-Echo in Steady-State (DESS) sequence, and biochemical images were obtained using a two-dimensional T2* mapping sequence. Cartilage score criteria ranged from 0 to 2 and were obtained using the Whole-Organ Magnetic Resonance Imaging Score (WORMS). The femoral, patellar, and tibial cartilages were automatically segmented and divided into subregions using the post-processing prototype software. Afterwards, all the subregions were carefully checked and manual corrections were done where needed. The dice coefficient correlations for each subregion by the automatic segmentation were calculated.ResultsCartilage volume after applying the manual correction was significantly lower than automatic segmentation (P < 0.05). The percentages of the cartilage volume change for each subregion after manual correction were all smaller than 5%. In all the subregions, the mean T2* relaxation time within manual corrected subregions was significantly lower than in regions after automatic segmentation (P < 0.05). The average time for the automatic segmentation of the whole knee was around 6 min, while the average time for manual correction of the whole knee was around 27 min.ConclusionsAutomatic segmentation of cartilage volume has a high dice coefficient correlation and it can provide accurate quantitative information about cartilage efficiently without individual bias.Advances in knowledge: Magnetic resonance imaging is the most promising method to detect structural changes in cartilage tissue. Unfortunately, due to the structure and morphology of the cartilages obtaining accurate segmentations can be problematic. There are some factors (location of cartilage subregions, hydrarthrosis and cartilage degeneration) that may influence segmentation accuracy. We therefore assessed the factors that influence segmentations error.

A Device-Independent, Shape Preserving Retinal Optical Coherence Tomography Image Alignment Method Applying TV-Unet for RPE Layer Detection

Retinal optical coherence tomography (OCT) images are widely used in diagnosis of ocular conditions. However, random shift and orientation changes of the retinal layers in OCT B-scans yield to appearance variations across the scans. These variations reduce the accuracy of the algorithms applied in the analysis of OCT images. In this study, we propose a preprocessing step to compensate these variations and align B-scans. At first, by incorporating total variation (TV) loss in the well-known Unet model, we propose a TV-Unet model to accurately detect the retinal pigment epithelium (RPE) layer in each B-scan. Then we use the detected RPE layer in the alignment method to form a curvature curve and a reference line. A novel window transferring-based alignment approach is applied to force the curve points to form a straight line, while preserving the shape and size of the pathological lesions. Since detection of RPE layer is a crucial step in the proposed alignment method, we utilized various datasets to train and test the TV-Unet and provided a multimodal, device-independent OCT image alignment method. The TV-Unet localizes the RPE layer in OCT images with low boundary error (maximum of 1.94pixels) and high Dice coefficient (minimum of 0.98). Quantitative and qualitative results indicated that the proposed method can efficiently detects the RPE layer and align OCT images while preserving the structure and size of the retinal lesions (biomarkers) in the OCT scans.