Project description:PurposeThe novel coronavirus COVID-19, which spread globally in late December 2019, is a global health crisis. Chest computed tomography (CT) has played a pivotal role in providing useful information for clinicians to detect COVID-19. However, segmenting COVID-19-infected regions from chest CT results is challenging. Therefore, it is desirable to develop an efficient tool for automated segmentation of COVID-19 lesions using chest CT. Hence, we aimed to propose 2D deep-learning algorithms to automatically segment COVID-19-infected regions from chest CT slices and evaluate their performance.Material and methodsHerein, 3 known deep learning networks: U-Net, U-Net++, and Res-Unet, were trained from scratch for automated segmenting of COVID-19 lesions using chest CT images. The dataset consists of 20 labelled COVID-19 chest CT volumes. A total of 2112 images were used. The dataset was split into 80% for training and validation and 20% for testing the proposed models. Segmentation performance was assessed using Dice similarity coefficient, average symmetric surface distance (ASSD), mean absolute error (MAE), sensitivity, specificity, and precision.ResultsAll proposed models achieved good performance for COVID-19 lesion segmentation. Compared with Res-Unet, the U-Net and U-Net++ models provided better results, with a mean Dice value of 85.0%. Compared with all models, U-Net gained the highest segmentation performance, with 86.0% sensitivity and 2.22 mm ASSD. The U-Net model obtained 1%, 2%, and 0.66 mm improvement over the Res-Unet model in the Dice, sensitivity, and ASSD, respectively. Compared with Res-Unet, U-Net++ achieved 1%, 2%, 0.1 mm, and 0.23 mm improvement in the Dice, sensitivity, ASSD, and MAE, respectively.ConclusionsOur data indicated that the proposed models achieve an average Dice value greater than 84.0%. Two-dimensional deep learning models were able to accurately segment COVID-19 lesions from chest CT images, assisting the radiologists in faster screening and quantification of the lesion regions for further treatment. Nevertheless, further studies will be required to evaluate the clinical performance and robustness of the proposed models for COVID-19 semantic segmentation.
Project description:We present a deep learning (DL)-based automated whole lung and COVID-19 pneumonia infectious lesions (COLI-Net) detection and segmentation from chest computed tomography (CT) images. This multicenter/multiscanner study involved 2368 (347'259 2D slices) and 190 (17 341 2D slices) volumetric CT exams along with their corresponding manual segmentation of lungs and lesions, respectively. All images were cropped, resized, and the intensity values clipped and normalized. A residual network with non-square Dice loss function built upon TensorFlow was employed. The accuracy of lung and COVID-19 lesions segmentation was evaluated on an external reverse transcription-polymerase chain reaction positive COVID-19 dataset (7'333 2D slices) collected at five different centers. To evaluate the segmentation performance, we calculated different quantitative metrics, including radiomic features. The mean Dice coefficients were 0.98 ± 0.011 (95% CI, 0.98-0.99) and 0.91 ± 0.038 (95% CI, 0.90-0.91) for lung and lesions segmentation, respectively. The mean relative Hounsfield unit differences were 0.03 ± 0.84% (95% CI, -0.12 to 0.18) and -0.18 ± 3.4% (95% CI, -0.8 to 0.44) for the lung and lesions, respectively. The relative volume difference for lung and lesions were 0.38 ± 1.2% (95% CI, 0.16-0.59) and 0.81 ± 6.6% (95% CI, -0.39 to 2), respectively. Most radiomic features had a mean relative error less than 5% with the highest mean relative error achieved for the lung for the range first-order feature (-6.95%) and least axis length shape feature (8.68%) for lesions. We developed an automated DL-guided three-dimensional whole lung and infected regions segmentation in COVID-19 patients to provide fast, consistent, robust, and human error immune framework for lung and pneumonia lesion detection and quantification.
Project description:Background Bone density measurement is an important examination for the diagnosis and screening of osteoporosis. The aim of this study was to develop a deep learning (DL) system for automatic measurement of bone mineral density (BMD) for osteoporosis screening using low-dose computed tomography (LDCT) images. Methods This retrospective study included 500 individuals who underwent LDCT scanning from April 2018 to July 2021. All images were manually annotated by a radiologist for the cancellous bone of target vertebrae and post-processed using quantitative computed tomography (QCT) software to identify osteoporosis. Patients were divided into the training, validation, and testing sets in a ratio of 6:2:2 using a 4-fold cross validation method. A localization model using faster region-based convolutional neural network (R-CNN) was trained to identify and locate the target vertebrae (T12–L2), then a 3-dimensional (3D) AnatomyNet was trained to finely segment the cancellous bone of target vertebrae in the localized image. A 3D DenseNet was applied for calculating BMD. The Dice coefficient was used to evaluate segmentation performance. Linear regression and Bland–Altman (BA) analyses were performed to compare the calculated BMD values using the proposed system with QCT. The diagnostic performance of the system for osteoporosis and osteopenia was evaluated with receiver operating characteristic (ROC) curve analysis. Results Our segmentation model achieved a mean Dice coefficient of 0.95, with Dice coefficients greater than 0.9 accounting for 96.6%. The correlation coefficient (R2) and mean errors between the proposed system and QCT in the testing set were 0.967 and 2.21 mg/cm3, respectively. The area under the curve (AUC) of the ROC was 0.984 for detecting osteoporosis and 0.993 for distinguishing abnormal BMD (osteopenia and osteoporosis). Conclusions The fully automated DL-based system is able to perform automatic BMD calculation for opportunistic osteoporosis screening with high accuracy using LDCT scans.
Project description:To evaluate diagnostic efficacy of deep learning (DL)-based automated bone mineral density (BMD) measurement for opportunistic screening of osteoporosis with routine computed tomography (CT) scans. A DL-based automated quantitative computed tomography (DL-QCT) solution was evaluated with 112 routine clinical CT scans from 84 patients who underwent either chest (N:39), lumbar spine (N:34), or abdominal CT (N:39) scan. The automated BMD measurements (DL-BMD) on L1 and L2 vertebral bodies from DL-QCT were validated with manual BMD (m-BMD) measurement from conventional asynchronous QCT using Pearson's correlation and intraclass correlation. Receiver operating characteristic curve (ROC) analysis identified the diagnostic ability of DL-BMD for low BMD and osteoporosis, determined by dual-energy X-ray absorptiometry (DXA) and m-BMD. Excellent concordance were seen between m-BMD and DL-BMD in total CT scans (r = 0.961/0.979). The ROC-derived AUC of DL-BMD compared to that of central DXA for the low-BMD and osteoporosis patients was 0.847 and 0.770 respectively. The sensitivity, specificity, and accuracy of DL-BMD compared to central DXA for low BMD were 75.0%, 75.0%, and 75.0%, respectively, and those for osteoporosis were 68.0%, 80.5%, and 77.7%. The AUC of DL-BMD compared to the m-BMD for low BMD and osteoporosis diagnosis were 0.990 and 0.943, respectively. The sensitivity, specificity, and accuracy of DL-BMD compared to m-BMD for low BMD were 95.5%, 93.5%, and 94.6%, and those for osteoporosis were 88.2%, 94.5%, and 92.9%, respectively. DL-BMD exhibited excellent agreement with m-BMD on L1 and L2 vertebrae in the various routine clinical CT scans and had comparable diagnostic performance for detecting the low-BMD and osteoporosis on conventional QCT.
Project description:ObjectiveDevelop two fully automatic osteoporosis screening systems using deep learning (DL) and radiomics (Rad) techniques based on low-dose chest CT (LDCT) images and evaluate their diagnostic effectiveness.MethodsIn total, 434 patients who underwent LDCT and bone mineral density (BMD) examination were retrospectively enrolled and divided into the development set (n = 333) and temporal validation set (n = 101). An automatic thoracic vertebra cancellous bone (TVCB) segmentation model was developed. The Dice similarity coefficient (DSC) was used to evaluate the segmentation performance. Furthermore, the three-class Rad and DL models were developed to distinguish osteoporosis, osteopenia, and normal bone mass. The diagnostic performance of these models was evaluated using the receiver operating characteristic (ROC) curve and decision curve analysis (DCA).ResultsThe automatic segmentation model achieved excellent segmentation performance, with a mean DSC of 0.96 ± 0.02 in the temporal validation set. The Rad model was used to identify osteoporosis, osteopenia, and normal BMD in the temporal validation set, with respective area under the receiver operating characteristic curve (AUC) values of 0.943, 0.801, and 0.932. The DL model achieved higher AUC values of 0.983, 0.906, and 0.969 for the same categories in the same validation set. The Delong test affirmed that both models performed similarly in BMD assessment. However, the accuracy of the DL model is 81.2%, which is better than the 73.3% accuracy of the Rad model in the temporal validation set. Additionally, DCA indicated that the DL model provided a greater net benefit compared to the Rad model across the majority of the reasonable threshold probabilities Conclusions: The automated segmentation framework we developed can accurately segment cancellous bone on low-dose chest CT images. These predictive models, which are based on deep learning and radiomics, provided comparable diagnostic performance in automatic BMD assessment. Nevertheless, it is important to highlight that the DL model demonstrates higher accuracy and precision than the Rad model.
Project description:ObjectiveWe aimed to develop a deep neural network for segmenting lung parenchyma with extensive pathological conditions on non-contrast chest computed tomography (CT) images.Materials and methodsThin-section non-contrast chest CT images from 203 patients (115 males, 88 females; age range, 31-89 years) between January 2017 and May 2017 were included in the study, of which 150 cases had extensive lung parenchymal disease involving more than 40% of the parenchymal area. Parenchymal diseases included interstitial lung disease (ILD), emphysema, nontuberculous mycobacterial lung disease, tuberculous destroyed lung, pneumonia, lung cancer, and other diseases. Five experienced radiologists manually drew the margin of the lungs, slice by slice, on CT images. The dataset used to develop the network consisted of 157 cases for training, 20 cases for development, and 26 cases for internal validation. Two-dimensional (2D) U-Net and three-dimensional (3D) U-Net models were used for the task. The network was trained to segment the lung parenchyma as a whole and segment the right and left lung separately. The University Hospitals of Geneva ILD dataset, which contained high-resolution CT images of ILD, was used for external validation.ResultsThe Dice similarity coefficients for internal validation were 99.6 ± 0.3% (2D U-Net whole lung model), 99.5 ± 0.3% (2D U-Net separate lung model), 99.4 ± 0.5% (3D U-Net whole lung model), and 99.4 ± 0.5% (3D U-Net separate lung model). The Dice similarity coefficients for the external validation dataset were 98.4 ± 1.0% (2D U-Net whole lung model) and 98.4 ± 1.0% (2D U-Net separate lung model). In 31 cases, where the extent of ILD was larger than 75% of the lung parenchymal area, the Dice similarity coefficients were 97.9 ± 1.3% (2D U-Net whole lung model) and 98.0 ± 1.2% (2D U-Net separate lung model).ConclusionThe deep neural network achieved excellent performance in automatically delineating the boundaries of lung parenchyma with extensive pathological conditions on non-contrast chest CT images.
Project description:BackgroundComputed tomography (CT) chest scans have become commonly used in clinical diagnosis. Image quality assessment (IQA) for CT images plays an important role in CT examination. It is worth noting that IQA is still a manual and subjective process, and even experienced radiologists make mistakes due to human limitations (fatigue, perceptual biases, and cognitive biases). There are also kinds of biases because of poor consensus among radiologists. Excellent IQA methods can reliably give an objective evaluation result and also reduce the workload of radiologists. This study proposes a deep learning (DL)-based automatic IQA method, to assess whether the image quality of respiratory phase on CT chest images are optimal or not, so that the CT chest images can be used in the patient's physical condition assessment.MethodsThis retrospective study analysed 212 patients' chest CT images, with 188 patients allocated to a training set (150 patients), validation set (18 patients), and a test set (20 patients). The remaining 24 patients were used for the observer study. Data augmentation methods were applied to address the problem of insufficient data. The DL-based IQA method combines image selection, tracheal carina segmentation, and bronchial beam detection. To automatically select the CT image containing the tracheal carina, an image selection model was employed. Afterward, the area-based approach and score-based approach were proposed and used to further optimize the tracheal carina segmentation and bronchial beam detection results, respectively. Finally, the score about the image quality of the patient's respiratory phase images given by the DL-based automatic IQA method was compared with the mean opinion score (MOS) given in the observer study, in which four blinded experienced radiologists took part.ResultsThe DL-based automatic IQA method achieved good performance in assessing the image quality of the respiratory phase images. For the CT sequence of the same patient, the DL-based IQA method had an accuracy of 92% in the assessment score, while the radiologists had an accuracy of 88%. The Kappa value of the assessment score between the DL-based IQA method and radiologists was 0.75, with a sensitivity of 85%, specificity of 91%, positive predictive value (PPV) of 92%, negative predictive value (NPV) of 93%, and accuracy of 88%.ConclusionsThis study develops and validates a DL-based automatic IQA method for the respiratory phase on CT chest images. The performance of this method surpassed that of the experienced radiologists on the independent test set used in this study. In clinical practice, it is possible to reduce the workload of radiologists and minimize errors caused by human limitations.
Project description:BackgroundLow-dose chest computed tomography (LDCT) screening improves early detection of lung cancer but poses challenges such as false positives and overdiagnosis, especially for nodules smaller than 8 mm where follow-up guidelines are unclear. Traditional risk prediction models have limitations, and deep learning (DL) algorithms offer potential improvements but often require large datasets. This study aimed to develop a DL-based, label-free lung cancer risk prediction model using alternative LDCT images and validate it in individuals without non-calcified solid pulmonary nodules larger than 8 mm.MethodsWe utilized LDCT scans from individuals without non-calcified solid nodules larger than 8 mm to develop a DL-based lung cancer risk prediction model. An alternative training dataset included 1,064 LDCT scans: 380 from patients with pathologically confirmed lung cancer and 684 from control individuals without lung cancer development over 5 years. For the lung cancer group, only the contralateral lung (without the tumor) was analyzed to represent high-risk individuals without large nodules. The LDCT scans were randomly divided into training and validation sets in a 3:1 ratio. Four three-dimensional (3D) convolutional neural networks (CNNs; 3D-CNN, MobileNet v2, SEResNet18, EfficientNet-B0) were trained using densely connected U-Net (DenseUNet)-segmented lung parenchyma images. The models were validated on a real-world test dataset comprising 1,306 LDCT scans (1,254 low-risk and 52 high-risk individuals) and evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), Brier scores, and calibration measures.ResultsIn the validation dataset, the AUC values were 0.801 for 3D-CNN, 0.802 for MobileNet v2, 0.755 for EfficientNet-B0, and 0.833 for SEResNet18. Corresponding Brier scores were 0.169, 0.175, 0.217, and 0.156, respectively, indicating good calibration, especially for SEResNet18. In the test dataset, the AUC values were 0.769 for 3D-CNN, 0.753 for MobileNet v2, 0.681 for EfficientNet-B0, and 0.820 for SEResNet18, with Brier scores of 0.169, 0.180, 0.202, and 0.138, respectively. The SEResNet18 model demonstrated the best performance, achieving the highest AUC and lowest Brier score in both validation and test datasets.ConclusionsOur study demonstrated that DL-based, label-free lung cancer risk prediction models using alternative LDCT images can effectively predict lung cancer development in individuals without non-calcified solid pulmonary nodules larger than 8 mm. By analyzing lung parenchyma on LDCT images without relying on nodule detection, these models may enhance the efficiency of LDCT screening programs. Further prospective studies are needed to determine their clinical utility and impact on screening protocols, and validation in larger, diverse populations is necessary to ensure generalizability.
Project description:Oxygen-induced retinopathy (OIR) is a widely used model to study ischemia-driven neovascularization (NV) in the retina and to serve in proof-of-concept studies in evaluating antiangiogenic drugs for ocular, as well as nonocular, diseases. The primary parameters that are analyzed in this mouse model include the percentage of retina with vaso-obliteration (VO) and NV areas. However, quantification of these two key variables comes with a great challenge due to the requirement of human experts to read the images. Human readers are costly, time-consuming, and subject to bias. Using recent advances in machine learning and computer vision, we trained deep learning neural networks using over a thousand segmentations to fully automate segmentation in OIR images. While determining the percentage area of VO, our algorithm achieved a similar range of correlation coefficients to that of expert inter-human correlation coefficients. In addition, our algorithm achieved a higher range of correlation coefficients compared with inter-expert correlation coefficients for quantification of the percentage area of neovascular tufts. In summary, we have created an open-source, fully automated pipeline for the quantification of key values of OIR images using deep learning neural networks.
Project description:PurposeThe deformable nature of the liver can make focal treatment challenging and is not adequately addressed with simple rigid registration techniques. More advanced registration techniques can take deformations into account (eg, biomechanical modeling) but require segmentations of the whole liver for each scan, which is a time-intensive process. We hypothesize that fully convolutional networks can be used to rapidly and accurately autosegment the liver, removing the temporal bottleneck for biomechanical modeling.Methods and materialsManual liver segmentations on computed tomography scans from 183 patients treated at our institution and 30 scans from the Medical Image Computing & Computer Assisted Intervention challenges were collected for this study. Three architectures were investigated for rapid automated segmentation of the liver (VGG-16, DeepLabv3 +, and a 3-dimensional UNet). Fifty-six cases were set aside as a final test set for quantitative model evaluation. Accuracy of the autosegmentations was assessed using Dice similarity coefficient and mean surface distance. Qualitative evaluation was also performed by 3 radiation oncologists on 50 independent cases with previously clinically treated liver contours.ResultsThe mean (minimum-maximum) mean surface distance for the test groups with the final model, DeepLabv3 +, were as follows: μContrast(N = 17): 0.99 mm (0.47-2.2), μNon_Contrast(N = 19)l: 1.12 mm (0.41-2.87), and μMiccai(N = 30)t: 1.48 mm (0.82-3.96). The qualitative evaluation showed that 30 of 50 autosegmentations (60%) were preferred to manual contours (majority voting) in a blinded comparison, and 48 of 50 autosegmentations (96%) were deemed clinically acceptable by at least 1 reviewing physician.ConclusionsThe autosegmentations were preferred compared with manually defined contours in the majority of cases. The ability to rapidly segment the liver with high accuracy achieved in this investigation has the potential to enable the efficient integration of biomechanical model-based registration into a clinical workflow.