Original CT Research Articles

Feasibility of synthetic computed tomography generated with an adversarial network for multi-sequence magnetic resonance-based brain radiotherapy.

ABSTRACTThe aim of this work is to generate synthetic computed tomography (sCT) images from multi-sequence magnetic resonance (MR) images using an adversarial network and to assess the feasibility of sCT-based treatment planning for brain radiotherapy. Datasets for 15 patients with glioblastoma were selected and 580 pairs of CT and MR images were used. T1-weighted, T2-weighted and fluid-attenuated inversion recovery MR sequences were combined to create a three-channel image as input data. A conditional generative adversarial network (cGAN) was trained using image patches. The image quality was evaluated using voxel-wise mean absolute errors (MAEs) of the CT number. For the dosimetric evaluation, 3D conformal radiotherapy (3D-CRT) and volumetric modulated arc therapy (VMAT) plans were generated using the original CT set and recalculated using the sCT images. The isocenter dose and dose–volume parameters were compared for 3D-CRT and VMAT plans, respectively. The equivalent path length was also compared. The mean MAEs for the whole body, soft tissue and bone region were 108.1 ± 24.0, 38.9 ± 10.7 and 366.2 ± 62.0 hounsfield unit, respectively. The dosimetric evaluation revealed no significant difference in the isocenter dose for 3D-CRT plans. The differences in the dose received by 2% of the volume (D2%), D50% and D98% relative to the prescribed dose were <1.0%. The overall equivalent path length was shorter than that for real CT by 0.6 ± 1.9 mm. A treatment planning study using generated sCT detected only small, clinically negligible differences. These findings demonstrated the feasibility of generating sCT images for MR-only radiotherapy from multi-sequence MR images using cGAN.

A deep residual learning network for predicting lung adenocarcinoma manifesting as ground-glass nodule on CT images.

To develop a deep learning-based artificial intelligence (AI) scheme for predicting the likelihood of the ground-glass nodule (GGN) detected on CT images being invasive adenocarcinoma (IA) and also compare the accuracy of this AI scheme with that of two radiologists. First, we retrospectively collected 828 histopathologically confirmed GGNs of 644 patients from two centers. Among them, 209 GGNs are confirmed IA and 619 are non-IA, including 409 adenocarcinomas in situ and 210 minimally invasive adenocarcinomas. Second, we applied a series of pre-preprocessing techniques, such as image resampling, rescaling and cropping, and data augmentation, to process original CT images and generate new training and testing images. Third, we built an AI scheme based on a deep convolutional neural network by using a residual learning architecture and batch normalization technique. Finally, we conducted an observer study and compared the prediction performance of the AI scheme with that of two radiologists using an independent dataset with 102 GGNs. The new AI scheme yielded an area under the receiver operating characteristic curve (AUC) of 0.92 ± 0.03 in classifying between IA and non-IA GGNs, which is equivalent to the senior radiologist's performance (AUC 0.92 ± 0.03) and higher than the score of the junior radiologist (AUC 0.90 ± 0.03). The Kappa value of two sets of subjective prediction scores generated by two radiologists is 0.6. The study result demonstrates using an AI scheme to improve the performance in predicting IA, which can help improve the development of a more effective personalized cancer treatment paradigm. • The feasibility of using a deep learning method to predict the likelihood of the ground-glass nodule being invasive adenocarcinoma. • Residual learning-based CNN model improves the performance in classifying between IA and non-IA nodules. • Artificial intelligence (AI) scheme yields higher performance than radiologists in predicting invasive adenocarcinoma.

Data augmentation using generative adversarial networks (CycleGAN) to improve generalizability in CT segmentation tasks

Labeled medical imaging data is scarce and expensive to generate. To achieve generalizable deep learning models large amounts of data are needed. Standard data augmentation is a method to increase generalizability and is routinely performed. Generative adversarial networks offer a novel method for data augmentation. We evaluate the use of CycleGAN for data augmentation in CT segmentation tasks. Using a large image database we trained a CycleGAN to transform contrast CT images into non-contrast images. We then used the trained CycleGAN to augment our training using these synthetic non-contrast images. We compared the segmentation performance of a U-Net trained on the original dataset compared to a U-Net trained on the combined dataset of original data and synthetic non-contrast images. We further evaluated the U-Net segmentation performance on two separate datasets: The original contrast CT dataset on which segmentations were created and a second dataset from a different hospital containing only non-contrast CTs. We refer to these 2 separate datasets as the in-distribution and out-of-distribution datasets, respectively. We show that in several CT segmentation tasks performance is improved significantly, especially in out-of-distribution (noncontrast CT) data. For example, when training the model with standard augmentation techniques, performance of segmentation of the kidneys on out-of-distribution non-contrast images was dramatically lower than for in-distribution data (Dice score of 0.09 vs. 0.94 for out-of-distribution vs. in-distribution data, respectively, p < 0.001). When the kidney model was trained with CycleGAN augmentation techniques, the out-of-distribution (non-contrast) performance increased dramatically (from a Dice score of 0.09 to 0.66, p < 0.001). Improvements for the liver and spleen were smaller, from 0.86 to 0.89 and 0.65 to 0.69, respectively. We believe this method will be valuable to medical imaging researchers to reduce manual segmentation effort and cost in CT imaging.

Lungs cancer classification from CT images: An integrated design of contrast based classical features fusion and selection

Lung cancer is a fatal type of cancer and it causes of severe deaths of approximately 422 people every day, worldwide. However, an early diagnosis is an expedient requirement for increasing the chances of human survival. In this regard the existing techniques of tumor detection, CT scans are mostly utilized to recognize the infected regions, nevertheless, the major challenges of CT images are low visibility of tumor regions, negative rates, to name but a few. In this work, we propose a novel design of contrast stretching based classical features fusion process for localizing the of lungs cancer classification. The proposed method encompasses three significant steps: firstly, contrast of original CT images is improved by gamma correction max intensity weights approach. Secondly, multiple texture, point, and geometric features are extracted from contrast images, which are later subjected to a serial canonical correlation-based fusion. Thirdly, zero values and negative features are replaced by an entropy-based approach, followed by weighted NCA for selection. Most discriminate features are fed into ensemble classifier for final classification. The validation of the proposed method is conducted on publicly available dataset: Lungs Data Science Bowl 2017 to achieving maximum accuracy of 99.4%. The numerical results clearly show that the performance of our proposed method outperforms in comparison with several existing methods with greater accuracy.

Preliminary value of CT radiomics in predicting anaplastic lymphoma kinase fusion gene expression in lung adenocarcinoma

Objective To explore the value of quantitative CT radiomics features in predicting the anaplastic lymphoma kinase (ALK) mutation status in lung adenocarcinoma patients. Methods This retrospective study reviewed one hundred and ninety-five lung adenocarcinoma patients (including 60 patients with ALK mutation) whose ALK genetic test results were available from Nov 2015 to May 2018 in PUMCH. VOIs were labeled by an automatic pulmonary nodule detection and segmentation algorithm and were later revised and confirmed by two senior radiologists. The PyRadiomics tools were used to resample the labeled regions, followed by image pre-processing (Wavelet filter or Laplacian of Gaussian (LoG) filter) and feature extraction. Normalized features were selected based on their representativeness on Dr. Wise research platform. Multivariate logistic regression was performed to develop prediction models of ALK mutation gene based on different image pre-processing techniques and different radiomics feature types. The results were validated by ten runs of five-fold cross validation. ROC curve analysis and Delong test were used to compare the predictive performance among models. Results Fifteen radiomics features with the highest representativeness were selected from the original 1 232 features. The prediction model based on these radiomics features showed good performance (AUC=0.88 in the training set and 0.78 in the validation set) and was not significantly different from the prediction models based on radiomics features of different pre-processing images (AUC=0.76, P=0.1, original CT images; AUC=0.75, P=0.3, Wavelet-filtered images; AUC=0.76, P=0.2, LoG-filtered images). Among the models built with radiomics features of different types, the one based on GLCM feature (a subtype of texture feature) showed the best performance in predicting ALK genetic status (AUC=0.83, accuracy=0.74, sensitivity=0.85 and specificity=0.69). The model based on first-order statistic features had an AUC of 0.80. Conclusion Quantitative CT radiomics features have a good potential to anticipate the expression of ALK fused gene in patients with lung adenocarcinoma. Key words: Lung neoplasm; Radiomics; Tomography, X-ray computed; Anaplastic lymphoma kinase

Evaluation of a deep learning-based pelvic synthetic CT generation technique for MRI-based prostate proton treatment planning

The purpose of this work is to validate the application of a deep learning-based method for pelvic synthetic CT (sCT) generation that can be used for prostate proton beam therapy treatment planning. We propose to integrate dense block minimization into 3D cycle-consistent generative adversarial networks (cycleGAN) framework to effectively learn the nonlinear mapping between MRI and CT pairs. A cohort of 17 patients with co-registered CT and MR pairs were used to test the deep learning-based sCT generation method by leave-one-out cross-validation. Image quality between the sCT and CT images, gamma analysis passing rate, dose-volume metrics, distal range displacement, and the individual pencil beam Bragg peak shift between sCT- and CT-based proton plans were evaluated. The average mean absolute error (MAE) was 51.32 ± 16.91 HU. The relative differences of the statistics of the PTV dose-volume histogram (DVH) metrics in between sCT and CT were generally less than 1%. Mean values of dose difference, absolute dose difference (in percent of the prescribed dose) were −0.07% ± 0.07% and 0.23% ± 0.08%. Mean gamma analysis pass rate of 1 mm/1%, 2 mm/2%, 3 mm/3% criteria with 10% dose threshold were 92.39% ± 5.97%, 97.95% ± 2.95% and 98.97% ± 1.62% respectively. The median, mean and standard deviation of absolute maximum range differences were 0.09 cm and 0.23 ± 0.25 cm. The median and mean Bragg peak shifts among the 17 patients were 0.09 cm and 0.18 ± 0.07 cm. The image similarity, dosimetric and distal range agreement between sCT and original CT suggests the feasibility of further development of an MRI-only workflow for prostate proton radiotherapy.

Prediction of Collapse Using Patient-Specific Finite Element Analysis of Osteonecrosis of the Femoral Head.

ObjectiveTo develop a prediction method for femoral head collapse by using patient‐specific finite element analysis of osteonecrosis of the femoral head (ONFH).MethodsThe retrospective study recruited 40 patients with ARCO stage‐II ONFH (40 pre‐collapse hips). Patients were divided into two groups according to the 1‐year follow‐up outcomes: patient group without femoral head collapse (noncollapse group, n = 20) and patient group with collapse (collapse group, n = 20). CT scans of the hip were performed for all patients once they joined the study. Patient‐specific finite element models were generated based on these original CT images following the same procedures: segmenting the necrotic lesion and viable proximal femur, meshing the computational models, assigning different material properties according to the Hounsfield unit distribution, simulating the stress loading of the slow walking gait, and measuring the distribution of the von Mises stress. Receiver operating characteristic (ROC) curve analysis was used to evaluate the predictive performance of the maximum level of the von Mises stress. The optimal cut‐off value was selected based on the Youden index and the corresponding predictive accuracy was reported as well.ResultsThe mean level of the maximum von Mises stress in the collapse group was 2.955 ± 0.539 MPa, whereas the mean stress level in the noncollapse group was 1.923 ± 0.793 MPa (P < 0.01). ROC analysis of the maximum von Mises stress found that the area under the ROC curve was 0.842 (95% CI: 0.717–0.968, P < 0.01). The maximum Youden index was 0.60, which corresponded to two optimal cut‐off values: 2.7801 MPa (sensitivity: 0.70; specificity: 0.90; predictive accuracy: 80.00%; LR+: 7), and 2.7027 MPa (sensitivity: 0.75; specificity: 0.85; predictive accuracy: 77.50%; LR+: 5).ConclusionFinite element analysis is a potential method for femoral head collapse prediction among pre‐collapse ONFH patients. The maximum level of the von Mises stress on the weight‐bearing surface of the femoral head could be a good biomechanical marker to classify the collapse risk. The collapse prediction method based on patient‐specific finite element analysis is, thus, suitable to apply to clinical practice, but further testing on a larger dataset is desirable.

Radiomics in predicting tumor molecular marker P63 for non-small cell lung cancer.

To establish a radiomics signature based on CT images of non-small cell lung cancer (NSCLC) to predict the expression of molecular marker P63. Methods: A total of 245 NSCLC patients who underwent CT scans were retrospectively included. All patients were confirmed by histopathological examinations and P63 expression were examined within 2 weeks after CT examination. Radiomics features were extracted by MaZda software and subjective image features were defined from original non-enhanced CT images. The Lasso-logistic regression model was used to select features and develop radiomics signature, subjective image features model, and combined diagnostic model. The predictive performance of each model was evaluated by the receiver operating characteristic (ROC) curve, and compared with Delong test. Results: Of the 245 patients, 96 were P63 positive and 149 were P63 negative. The subjective image feature model consisted of 6 image features. Through feature selection, the radiomics signature consisted of 8 radiomics features. The area under the ROC curves of the subjective image feature model and the radiomics signature in predicting P63 expression statue were 0.700 and 0.755, respectively, without a significant difference (P>0.05). The combined diagnostic model showed the best predictive power (AUC=0.817, P<0.01). Conclusion: The radiomics-based CT scan images can predict the expression status of NSCLC molecular marker P63. The combination of the radiomics features and subjective image features can significantly improve the predictive performance of the predictive model, which may be helpful to provide a non-invasive way for understanding the molecular information for lung cancer cells.

Evaluation of Intensity- and Contour-Based Deformable Image Registration Accuracy in Pancreatic Cancer Patients.

We aimed to clarify the accuracy of rigid image registration and deformable image registration (DIR) in carbon-ion radiotherapy (CIRT) for pancreatic cancer. Six patients with pancreatic cancer who were treated with passive irradiation CIRT were enrolled. Three registration patterns were evaluated: treatment planning computed tomography images (TPCT) to CT images acquired in the treatment room (IRCT) in the supine position, TPCT to IRCT in the prone position, and TPCT in the supine position to the prone position. After warping the contours of the original CT images to the destination CT images using deformation matrices from the registration, the warped delineated contours on the destination CT images were compared with the original ones using mean displacement to agreement (MDA). Four contours (clinical target volume (CTV), gross tumor volume (GTV), stomach, duodenum) and four registration algorithms (rigid image registration [RIR], intensity-based DIR [iDIR], contour-based DIR [cDIR], and a hybrid iDIR-cDIR ([hDIR]) were evaluated. The means ± standard deviation of the MDAs of all contours for RIR, iDIR, cDIR, and hDIR were 3.40 ± 3.30, 2.2 1± 2.48, 1.46 ± 1.49, and 1.46 ± 1.37 mm, respectively. There were significant differences between RIR and iDIR, and between RIR/iDIR and cDIR/hDIR. For the pancreatic cancer patient images, cDIR and hDIR had better accuracy than RIR and iDIR.

Risk of Ischemic Heart Disease in Patients After Conccurent Chemoradiotherapy for Lung and Esophageal Cancer: A Case-Control Study

Late cardiac toxicity after thoracic radiotherapy is an important problem because the number of long-term survivors are increasing due to advances in medicine. However, little is known about the effect of incidental exposure of the heart to radiation on the risk of subsequent ischemic heart disease. Purpose of this study was to investigate the relationship between cardiac radiation dose and major adverse cardiovascular event (MACE). We conducted a case-control study of MACE in patients who underwent definitive thoracic radiotherapy for lung or esophageal cancer between September 2002 to December 2005 at our institution. MACE were defined as myocardial infarction, coronary revascularization, or death from ischemic heart disease. To minimize selection bias between MACE and non-MACE group, a propensity score matching was used based on the following covariates: age, sex, smoking history, and history of ischemic heart disease. The mean heart dose (MHD) was recalculated using original volumetric CT simulation images and radiotherapy planning system. Beam arrangements, energy, and monitor units were simulated based on the patient’s radiotherapy record. Among 543 patients, 14 patients developed MACE and 63 patients were included in the matched non-MACE cohort. Forty-seven and 30 patients had lung and esophageal cancer, respectively. All patients underwent three dimensional conformal radiotherapy with concurrent chemotherapy. Mean prescribed dose was 56 Gy. Median follow up time was 114.2 months. There was no statistically significant difference in overall survival rate between two groups. We set the cutoff value of 30 Gy in MHD by receiver operating characteristic curve analysis to predict MACE. There was a significant difference in the incident of MACE between high-MHD (>=30Gy) and low-MHD (<30Gy) groups (10.0 vs 47.1%, Fisher’s exact test, p=0.002). Patients in low-MHD group had a longer MACE-free survival than those in high-MHD group (94.8 months vs not-reached, Log-rank test p=0.044). Exposure of the heart to ionizing radiation in thoracic radiotherapy may increase the subsequent rate of ischemic heart disease.

An investigation of CNN models for differentiating malignant from benign lesions using small pathologically proven datasets

Cancer has been one of the most threatening diseases to human health. There have been many efforts devoted to the advancement of radiology and transformative tools (e.g. non-invasive computed tomographic or CT imaging) to detect cancer in early stages. One of the major goals is to identify malignant from benign lesions. In recent years, machine deep learning (DL), e.g. convolutional neural network (CNN), has shown encouraging classification performance on medical images. However, DL algorithms always need large datasets with ground truth. Yet in the medical imaging field, especially for cancer imaging, it is difficult to collect such large volume of images with pathological information. Therefore, strategies are needed to learn effectively from small datasets via CNN models. To forward that goal, this paper explores two CNN models by focusing extensively on expansion of training samples from two small pathologically proven datasets (colorectal polyp dataset and lung nodule dataset) and then differentiating malignant from benign lesions. Experimental outcomes indicate that even in very small datasets of less than 70 subjects, malignance can be successfully differentiated from benign via the proposed CNN models, the average AUCs (area under the receiver operating curve) of differentiating colorectal polyps and pulmonary nodules are 0.86 and 0.71, respectively. Our experiments further demonstrate that for these two small datasets, instead of only studying the original raw CT images, feeding additional image features, such as the local binary pattern of the lesions, into the CNN models can significantly improve classification performance. In addition, we find that our explored voxel level CNN model has better performance when facing the small and unbalanced datasets.

Personalized 3D printed coronary models in coronary stenting

3D printing has shown great promise in cardiovascular disease, with reports mainly focusing on pre-surgical planning and medical education. Research on utilization of 3D printed models in simulating coronary stenting has not been reported. In this study, we presented our experience of placing coronary stents into personalized 3D printed coronary models with the aim of determining stent lumen visibility with images reconstructed with different postprocessing views and algorithms. A total of six coronary stents with diameter ranging from 2.5 to 4.0 mm were placed into 3 patient-specific 3D printed coronary models for simulation of coronary stenting. The 3D printed models were placed in a plastic container and scanned on a 192-slice third generation dual-source CT scanner with images reconstructed with soft (Bv36) and sharp (Bv59) kernel algorithms. Thick and thin slab maximum-intensity projection (MIP) images were also generated from the original CT data for comparison of stent lumen visibility. Stent lumen diameter was measured on 2D axial and MIP images, while stent diameter was measured on 3D volume rendering images. 3D virtual intravascular endoscopy (VIE) images were generated to provide intraluminal views of the coronary wall and stent appearances. All of these stents were successfully placed into the right and left coronary arteries but 2 of them did not obtain wall apposition along the complete length. The stent lumen visibility ranged from 54 to 97%, depending on the stent location in the coronary arteries. The mean stent lumen diameters measured on 2D axial, thin and thick slab MIP images were found to be significantly smaller than the actual size (P<0.01). Thick slab MIP images resulted in measured stent lumen diameters smaller than those from thin slab MIP images, with significant differences noticed in most of the measurements (4 out of 6 stents) (P<0.05), and no significant differences in the remaining 2 stents (P=0.19-0.38). In contrast, 3D volume rendering images allowed for more accurate measurements with measured stent diameters close to the actual dimensions in most of these coronary stents, except for the stent placed at the right coronary artery in one of the models due to insufficient expansion of the stent. Images reconstructed with sharp kernel Bv59 significantly improved stent lumen visibility when compared to the smooth Bv36 kernel (P=0.01). 3D VIE was successfully generated in all of the datasets with clear visualization of intraluminal views of the stents in relation to the coronary wall. This preliminary report shows the feasibility of using 3D printed coronary artery models in coronary stenting for investigation of optimal coronary CT angiography protocols. Future studies should focus on placement of more stents with a range of stent diameters in the quest to reduce the need for invasive angiography for surveillance.

Normal tissue complication probability models in plan evaluation of children with brain tumors referred to proton therapy

Background: Children with brain tumors undergoing radiotherapy are at particular risk of radiation-induced morbidity and are therefore routinely considered for proton therapy (PT) to reduce the dose to healthy tissues. The aim of this study was to apply pediatric constraints and normal tissue complication probability (NTCP) models when evaluating the differences between PT and contemporary photon-based radiotherapy, volumetric modulated arc therapy (VMAT).Methods: Forty patients (aged 1–17 years) referred from Norwegian institutions to cranial PT abroad during 2014–2016 were selected for VMAT re-planning using the original CT sets and target volumes. The VMAT and delivered PT plans were compared by dose/volume metrics and NTCP models related to growth hormone deficiency, auditory toxicity, visual impairment, xerostomia, neurocognitive outcome and secondary brain and parotid gland cancers.Results: The supratentorial brain, temporal lobes, hippocampi, hypothalamus, pituitary glands, cochleas, salivary glands, optic nerves and chiasm received lower mean doses from PT. Reductions in population median NTCP were significant for auditory toxicity (VMAT: 3.8%; PT: 0.3%), neurocognitive outcome (VMAT: 3.0 IQ points decline at 5 years post RT; PT: 2.5 IQ points), xerostomia (VMAT: 2.0%; PT: 0.6%), excess absolute risk of secondary cancer of the brain (VMAT: 9.2%; PT: 6.7%) and salivary glands (VMAT: 2.8%; PT:0.5%). Across all patients, 23/38 PT plans had better or comparable estimated risks for all endpoints (within ±10% of the risk relative to VMAT), whereas for 1/38 patients all estimates were better or comparable with VMAT.Conclusions: PT reduced the volumes of normal tissues exposed to radiation, particularly low-to-intermediate dose levels, and this was reflected in lower NTCP. Of the included endpoints, substantial reductions in population medians were seen from the delivered PT plans for auditory complications, xerostomia, and risk of secondary cancers of the brain and salivary glands.

Volume Visualization for Improving CT Lung Nodule Detection.

Inspired by the outstanding performance of deep convolutional neural networks (CNNs), nowadays modern computer-aided detection (CAD) systems for CT lung nodules generally delve into 2D or 3D CNNs directly without considering traditional image preprocessing techniques. However, detection of large pulmonary nodules and masses are computationally challenging, especially for 3D CNNs. In this paper, we examine the possibility of using volume visualized CT thin-slab images with 2D CNNs to reduce computation complexity and improve CAD performance. We tested 4 types of images: original 2D CT, 2D projection of thin slabs, mixture by arranging original and projection in different color channels, and mixture by the pixelwise maximum intensity of original CT and projection. We evaluated these images on a dataset of 30 CT scans with 30 different-sized nodules and masses on GoogLeNet via a transfer learning and cross validation paradigm. We found that projection visualization alone had a better or equal area-under curve score for all the different-sized nodules and masses. However, mixture by the maximum of CT and projection demonstrated a preferred performance with a true positive rate of 0.8 and a false positive rate of 0.046 in detecting large nodules and masses.

Adding Depth to Cephalometric Analysis: Comparing Two- and Three-Dimensional Angular Cephalometric Measurements.

Lateral cephalometric radiographs (LCR) have been the standard tool used for cephalometric analysis in craniofacial surgery. Over the past decade, a three-dimensional (3D) revolution in cephalometric analysis and surgical planning has been underway. To date, research has not validated whether cephalometric measurements taken from two-dimensional (2D) and 3D data sources are equivalent and interchangeable. The authors sought to compare angular cephalometric measurements taken with 2D and 3D modalities. Sixty-two head CT scans (36 females, 26 males) with an average age of 63 ± 20 years were studied. Twelve cephalometric angular measurements were taken from 3D reconstructed skulls using the software package Mimics 19.0 (Materialize; Leuven, Belgium). These same facial angles were measured from 2D lateral cephalograms reconstructed from the original CT scans using Dolphin 11.9. Measurements taken with both techniques were compared for agreement using a paired t test. Intra-class correlation coefficient assessment was used to determine inter-rater reliability. Statistical significance was set at P < 0.05. Five of the 12 angular measurements (SNA, SNB, MP-FH, U1-SN, and U1-L1) demonstrated statistically significant differences (P < 0.05) between the 2D and 3D analyses. All of these differences were less than the standard deviations for the respective measure. The differences between angular cephalometric values obtained from 2D LCRs and 3D CT reconstructions are small. This supports the practices of using 2D and 3D cephalometric data interchangeably in most applications. Clinicians must be selective in which measures they employ to maximize accuracy and care must be taken when measuring dental inclination with lateral cephalograms.

MRI-based treatment planning for liver stereotactic body radiotherapy: validation of a deep learning-based synthetic CT generation method.

The purpose of this work is to develop and validate a learning-based method to derive electron density from routine anatomical MRI for potential MRI-based SBRT treatment planning. We proposed to integrate dense block into cycle generative adversarial network (GAN) to effectively capture the relationship between the CT and MRI for CT synthesis. A cohort of 21 patients with co-registered CT and MR pairs were used to evaluate our proposed method by the leave-one-out cross-validation. Mean absolute error, peak signal-to-noise ratio and normalized cross-correlation were used to quantify the imaging differences between the synthetic CT (sCT) and CT. The accuracy of Hounsfield unit (HU) values in sCT for dose calculation was evaluated by comparing the dose distribution in sCT-based and CT-based treatment planning. Clinically relevant dose-volume histogram metrics were then extracted from the sCT-based and CT-based plans for quantitative comparison. The mean absolute error, peak signal-to-noise ratio and normalized cross-correlation of the sCT were 72.87 ± 18.16 HU, 22.65 ± 3.63 dB and 0.92 ± 0.04, respectively. No significant differences were observed in the majority of the planning target volume and organ at risk dose-volume histogram metrics ( p > 0.05). The average pass rate of γ analysis was over 99% with 1%/1 mm acceptance criteria on the coronal plane that intersects with isocenter. The image similarity and dosimetric agreement between sCT and original CT warrant further development of an MRI-only workflow for liver stereotactic body radiation therapy. This work is the first deep-learning-based approach to generating abdominal sCT through dense-cycle-GAN. This method can successfully generate the small bony structures such as the rib bones and is able to predict the HU values for dose calculation with comparable accuracy to reference CT images.

MRI-only based synthetic CT generation using dense cycle consistent generative adversarial networks.

Automated synthetic computed tomography (sCT) generation based on magnetic resonance imaging (MRI) images would allow for MRI-only based treatment planning in radiation therapy, eliminating the need for CT simulation and simplifying the patient treatment workflow. In this work, the authors propose a novel method for generation of sCT based on dense cycle-consistent generative adversarial networks (cycle GAN), a deep-learning based model that trains two transformation mappings (MRI to CT and CT to MRI) simultaneously. The cycle GAN-based model was developed to generate sCT images in a patch-based framework. Cycle GAN was applied to this problem because it includes an inverse transformation from CT to MRI, which helps constrain the model to learn a one-to-one mapping. Dense block-based networks were used to construct generator of cycle GAN. The network weights and variables were optimized via a gradient difference (GD) loss and a novel distance loss metric between sCT and original CT. Leave-one-out cross-validation was performed to validate the proposed model. The mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and normalized cross correlation (NCC) indexes were used to quantify the differences between the sCT and original planning CT images. For the proposed method, the mean MAE between sCT and CT were 55.7 Hounsfield units (HU) for 24 brain cancer patients and 50.8HU for 20 prostate cancer patients. The mean PSNR and NCC were 26.6dB and 0.963 in the brain cases, and 24.5dB and 0.929 in the pelvis. We developed and validated a novel learning-based approach to generate CT images from routine MRIs based on dense cycle GAN model to effectively capture the relationship between the CT and MRIs. The proposed method can generate robust, high-quality sCT in minutes. The proposed method offers strong potential for supporting near real-time MRI-only treatment planning in the brain and pelvis.

Three-dimensional reconstruction of human kidney based on UroMedix-3D system and its application in kidney surgery

To explore the feasibility of rapid and accurate three-dimensional (3D) image reconstruction using Uromedix-3D software for urological surgery. The original renal thin-slice enhancement CT data were obtained from patients with kidney lesions treated in our hospital between December, 2015 and October, 2018. The self-developed Uromedix- 3D system was used to reconstruct the normal kidney structures, blood vessels, collecting systems and the lesions. The spatial anatomic relationships of the structures were measured and digitized for surgical planning. 3D reconstruction of the kidneys was performed in a total of 173 cases, and the mean time for reconstruction was 31.24±2.012 min. Of these cases, 147 (84.9%) had renal tumors, and 2 had renal tumors with tumor thrombus. In addition to renal tumors, the Uromedix-3D system was also used for reconstructing other lesions including UPJO, kidney stones and retroperitoneal masses. Renal artery reconstruction was performed in 170 cases, which allowed observation of the precise terminal branches (up to 7th grade arterial branch) of the artery; 109 (64%) cases showed the 5th grade arterial branch or above. Renal artery variations were detected in 37 cases, including accessory renal artery (24 cases) and multiple renal arteries (13 cases). The renal veins were reconstructed in 164 cases, and second grade or above (up to the 4th grade) vein branches were observed in 138 (84%) cases. Uromedix-3D system can accurately and efficiently reconstruct the 3D structure of human kidneys and the renal lesions based on enhanced CT data. The reconstructed 3D model allows objective assessment of the spatial anatomical relationship of the lesions to provide assistance in surgical planning.

A fully automatic computer-aided diagnosis system for hepatocellular carcinoma using convolutional neural networks

The cancer of liver, which is the leading cause of cancer death, is commonly diagnosed by comparing the changes of gray level of liver tissue in the different phases of the patient's CT images. To aid the doctor in reducing misdiagnosis or missed diagnosis, a fully automatic computer-aided diagnosis (CAD) system is proposed to diagnose hepatocellular carcinoma (HCC) using convolutional neural network (CNN) classifier. The automatic segmentation and classification are two core technologies of the proposed CAD system, which are both realized based on CNN. The segmentation of liver and tumor is implemented by a fully convolutional networks (FCN) based on a fine tuning VGG-16 model with two additional ‘skip structures’ using a weighted loss function which helps to solve the problem of inaccurate tumor segmentation caused by the inevitably unbalanced training data. HCC classification is implemented by a 9-layer CNN classifier, whose input is a 4-channel image data constructed by combining the segmentation result of FCN with the original CT image. A total of 165 venous phase CT images including 46 diffuse tumors, 43 nodular tumors, and 76 massive tumors are used to evaluate the performance of the proposed CAD system. The classification accuracy of CNN classifier for diffuse, nodular and massive tumors are 98.4%, 99.7% and 98.7% respectively, which are significantly improved in contrast with the traditional feature-based ANN and SVM classifiers. The proposed CAD system, which is unaffected by the difference of preprocessing method and feature type, is proved satisfactory and feasible by the test set.

CT detection of primary and metastatic ileal carcinoid tumor: rates of missed findings and associated delay in clinical diagnosis.

To determine the rate of missed CT findings of ileal carcinoid tumor prior to pathologic diagnosis and the resultant diagnostic delay. Initially, 74 patients with abdominal and pelvic CT prior to pathologically-proven diagnosis of ileal carcinoid were identified. Patients were excluded when the original CT study (n = 6) or report (n = 4) was not available, resulting in a final cohort of 64 patients (mean age, 58.3years; 29M/35F); 27 (42%) patients had more than one abdominal CT prior to diagnosis. All available CT studies prior to diagnosis were retrospectively reviewed for the presence of the primary ileal tumor and metastatic disease (mesenteric and hepatic). Primary ileal tumors were prospectively missed on at least one CT scan in 64% (32/50) of patients with retrospectively identifiable disease. CT findings of mesenteric spread were missed at least once in 46% (25/54) of cases where present in retrospect. By the final pre-operative CT, hepatic metastases and bowel wall thickening were present in 55% (35/64) and 52% (33/64) of cases, respectively. In patients with missed ileal and/or mesenteric findings resulting in diagnostic delay, mean delay was 40months (range 4-98months). Initial presentation of ileal carcinoid tumor, even with mesenteric involvement, is often missed prospectively at abdominal CT, leading to delay in diagnosis until bowel or mesenteric findings become more obvious, or hepatic metastatic disease manifests. Radiologists should make a concerted effort to evaluate the bowel and mesentery in patients with long-standing vague abdominal symptoms.