Assessment of Pelvic Tilt in Anteroposterior Radiographs by Area Ratio Based on Deep Learning.

STUDY DESIGN: Diagnostics. OBJECTIVES: Based on deep learning semantic segmentation model, we sought to assess pelvic tilt by area ratio of the lesser pelvic and the obturator foramen in anteroposterior (AP) radiographs. BACKGROUND: Pelvic tilt is a critical factor in hip and spinal surgery, commonly evaluated by medical professionals through sagittal pelvic radiographs. The inherent pelvic asymmetry, as well as potential obstructions from clothing and musculature in roentgenography, may result in ghosting and blurring artifacts, thereby complicating precise measurement. METHODS: PT directly affects the area ratio of the lesser pelvis to the obturator foramen in AP radiographs. An exponential regression analysis of simulated radiographs from ten male and ten female pelvises in specific tilt positions derived a formula correlating this area ratio with PT. Two blinded investigators evaluated this formula using 161 simulated AP pelvic radiographs. A deep learning semantic segmentation model was then fine-tuned to automatically calculate the area ratio, enabling intelligent PT evaluation. This model and the regression function were integrated for automated PT measurement and tested on a dataset of 231 clinical cases. RESULTS: We observed no disparity between males and females in the aforementioned area ratio. The test results from two blinded investigators analyzing 161 simulated radiographs revealed a mean absolute error of 0.19° (SD±4.71°), with a correlation coefficient between them reaching 0.96. Additionally, the mean absolute error obtained from testing 231 clinical AP radiographs using the fine-tuned semantic segmentation model mentioned earlier is -0.58° (SD±5.97°). CONCLUSION: We found that using deep learning neural networks enabled a more accurate and robust automatic measurement of PT through the area ratio of the lesser pelvis and obturator foramen.

A lightweight super-resolution network with skip-connections.

INTRODUCTION: In some hospitals in remote areas, due to the lack of MRI scanners with high magnetic field intensity, only low-resolution MRI images can be obtained, hindering doctors from making correct diagnoses. In our study, higher-resolution images were obtained through low-resolution MRI images. Moreover, as our algorithm is a lightweight algorithm with a small number of parameters, it can be carried out in remote areas under the condition of the lack of computing resources. Moreover, our algorithm is of great clinical significance in providing references for doctors' diagnoses and treatment in remote areas. METHODS: We compared different super-resolution algorithms to obtain high-resolution MRI images, including SRGAN, SPSR, and LESRCNN. A global skip connection was applied to the original network of LESRCNN to use global semantic information to get better performance. RESULTS: Experiments reported that our network improved SSMI by 0.8% and also achieved an obvious increase in PSNR, PI, and LPIPS compared to LESRCNN in our dataset. Similar to LESRCNN, our network has a very short running time, a small number of parameters, low time complexity, and low space complexity while ensuring high performance compared to SRGAN and SPSR. Five MRI doctors were invited for a subjective evaluation of our algorithm. All agreed on significant improvements and that our algorithm could be used clinically in remote areas and has great value. CONCLUSION: The experimental results demonstrated the performance of our algorithm in super-resolution MRI image reconstruction. It allows us to obtain high-resolution images in the absence of high-field intensity MRI scanners, which have great clinical significance. The short running time, a small number of parameters, low time complexity, and low space complexity ensure that our network can be used in grassroots hospitals in remote areas that lack computing resources. We can reconstruct high-resolution MRI images in a short time, thus saving time for patients. Our algorithm can be biased towards practical applications; however, doctors have affirmed the clinical value of our algorithm.

AI-assisted Method for Efficiently Generating Breast Ultrasound Screening Reports.

BACKGROUND: Ultrasound is one of the preferred choices for early screening of dense breast cancer. Clinically, doctors have to manually write the screening report, which is time-consuming and laborious, and it is easy to miss and miswrite. AIM: We proposed a new pipeline to automatically generate AI breast ultrasound screening reports based on ultrasound images, aiming to assist doctors in improving the efficiency of clinical screening and reducing repetitive report writing. METHODS: AI efficiently generated personalized breast ultrasound screening preliminary reports, especially for benign and normal cases, which account for the majority. Doctors then make simple adjustments or corrections based on the preliminary AI report to generate the final report quickly. The approach has been trained and tested using a database of 4809 breast tumor instances. RESULTS: Experimental results indicate that this pipeline improves doctors' work efficiency by up to 90%, greatly reducing repetitive work. CONCLUSION: Personalized report generation is more widely recognized by doctors in clinical practice than non-intelligent reports based on fixed templates or options to fill in the blanks.

Cardiac Segmentation Method Based on Domain Knowledge.

Echocardiography plays an important role in the clinical diagnosis of cardiovascular diseases. Cardiac function assessment by echocardiography is a crucial process in daily cardiology. However, cardiac segmentation in echocardiography is a challenging task due to shadows and speckle noise. The traditional manual segmentation method is a time-consuming process and limited by inter-observer variability. In this paper, we present a fast and accurate echocardiographic automatic segmentation framework based on Convolutional neural networks (CNN). We propose FAUet, a segmentation method serially integrated U-Net with coordinate attention mechanism and domain feature loss from VGG19 pre-trained on the ImageNet dataset. The coordinate attention mechanism can capture long-range dependencies along one spatial direction and meanwhile preserve precise positional information along the other spatial direction. And the domain feature loss is more concerned with the topology of cardiac structures by exploiting their higher-level features. In this research, we use a two-dimensional echocardiogram (2DE) of 88 patients from two devices, Philips Epiq 7C and Mindray Resona 7T, to segment the left ventricle (LV), interventricular septal (IVS), and posterior left ventricular wall (PLVW). We also draw the gradient weighted class activation mapping (Grad-CAM) to improve the interpretability of the segmentation results. Compared with the traditional U-Net, the proposed segmentation method shows better performance. The mean Dice Score Coefficient (Dice) of LV, IVS, and PLVW of FAUet can achieve 0.932, 0.848, and 0.868, and the average Dice of the three objects can achieve 0.883. Statistical analysis showed that there is no significant difference between the segmentation results of the two devices. The proposed method can realize fast and accurate segmentation of 2DE with a low time cost. Combining coordinate attention module and feature loss with the original U-Net framework can significantly increase the performance of the algorithm.

Improved FCOS for Detecting Breast Cancers.

PURPOSE: Breast cancer ranks first among cancers affecting women's health. Our goal is to develop a fast, high-precision, and fully automated breast cancer detection algorithm to improve the early detection rate of breast cancer. METHODS: We compare different object detection algorithms, including anchor-based and anchor-free object detection algorithms for detecting breast lesions. Finally, we find that the fully convolutional onestage object detection (FCOS) showed the best performance in the detection of breast lesions, which is an anchor-free algorithm. 1) Considering that the detection of breast lesions requires the context information of the ultrasound images, we introduce the non-local technique, which models long-range dependency between pixels to the FCOS algorithm, providing the global context information for the detection of the breast lesions. 2) The variety of shapes and sizes of breast lesions makes detection difficult. We propose a new deformable spatial attention (DSA) module and add it to the FCOS algorithm. RESULTS: The detection performance of the original FCOS is that the average precision (AP) for benign lesions is 0.818, and for malignant lesions is 0.888. The FCOS with a non-local module improves the performance of the breast detection; the AP of benign lesions was 0.819, and that of malignant lesions was 0.894. Combining the DSA module with the FCOS improves the performance of breast detection; the AP for benign lesions and malignant lesions is 0.840 and 0.899, respectively. CONCLUSION: We propose two methods to improve the FCOS algorithm from different perspectives to improve its performance in detecting breast lesions. We find that FCOS combined with DSA is beneficial in improving the localization and classification of breast tumors and can provide auxiliary diagnostic advice for ultrasound physicians, which has a certain clinical application value.

Artificial Intelligence Medical Ultrasound Equipment: Application of Breast Lesions Detection.

Breast cancer ranks first among cancers affecting women's health. Our work aims to realize the intelligence of the medical ultrasound equipment with limited computational capability, which is used for the assistant detection of breast lesions. We embed the high-computational deep learning algorithm into the medical ultrasound equipment with limited computational capability by two techniques: (1) lightweight neural network: considering the limited computational capability of ultrasound equipment, a lightweight neural network is designed, which greatly reduces the amount of calculation. And we use the technique of knowledge distillation to train the low-precision network helped with the high-precision network; (2) asynchronous calculations: consider four frames of ultrasound images as a group; the image of the first frame of each group is used as the input of the network, and the result is respectively fused with the images of the fourth to seventh frames. An amount of computation of 30 GFLO/frame is required for the proposed lightweight neural network, about 1/6 of that of the large high-precision network. After trained from scratch using the knowledge distillation technique, the detection performance of the lightweight neural network (sensitivity = 89.25%, specificity = 96.33%, the average precision [AP] = 0.85) is close to that of the high-precision network (sensitivity = 98.3%, specificity = 88.33%, AP = 0.91). By asynchronous calculation, we achieve real-time automatic detection of 24 fps (frames per second) on the ultrasound equipment. Our work proposes a method to realize the intelligence of the low-computation-power ultrasonic equipment, and successfully achieves the real-time assistant detection of breast lesions. The significance of the study is as follows: (1) The proposed method is of practical significance in assisting doctors to detect breast lesions; (2) our method provides some practical and theoretical support for the development and engineering of intelligent equipment based on artificial intelligence algorithms.

Ultrasonic Diagnosis of Breast Nodules Using Modified Faster R-CNN.

Breast cancer has become the biggest threat to female health. Ultrasonic diagnosis of breast cancer based on artificial intelligence is basically a classification of benign and malignant tumors, which does not meet clinical demand. Besides, the current target detection method performs poorly in detecting small lesions, while it is clinically required to detect nodules below 2 mm. The objective of this study is to (a) propose a diagnostic method based on Breast Imaging Reporting and Data System (BI-RADS) and (b) increase its detectability of small lesions. We modified the framework of Faster R-CNN (Faster Region-based Convolutional Neural Network) by introducing multi-scale feature extraction and multi-resolution candidate bound extraction into the network. Then, it was trained using 852 images of BI-RADS C2, 739 images of C3, and 1662 images of malignancy (BI-RADS 4a/4b/4c/5/6). We compared our model with unmodified Faster R-CNN and YOLO v3 (You Only Look Once v3). The mean average precision (mAP) is significantly increased to 0.913, while its average detection speed is slightly declined to 4.11 FPS (frames per second). Meanwhile, its detectivity of small lesions is effectively improved. Moreover, we also tentatively applied our model on video sequences and got satisfactory results. We modified Faster R-CNN and trained it partly based on BI-RADS. Its detectability of lesions, as well as small nodules, was significantly improved. In view of wide coverage of dataset and satisfactory test results, our method can basically meet clinical needs.

Accurate and rapid CT image segmentation of the eyes and surrounding organs for precise radiotherapy.

OBJECTIVE: The precise segmentation of organs at risk (OARs) is of importance for improving therapeutic outcomes and reducing injuries of patients undergoing radiotherapy. In this study, we developed a new approach for accurate computed tomography (CT) image segmentation of the eyes and surrounding organs, which is first locating then segmentation (FLTS). METHODS: The FLTS approach was composed of two steps: (a) classification of CT images using convolutional neural networks (CNN), and (b) segmentation of the eyes and surrounding organs using modified U-shape networks. In order to obtain optimal performance, we enhanced our training datasets by random jitter and rotation. RESULTS: This model was trained and verified using the clinical datasets that were delineated by experienced physicians. The dice similarity coefficient (DSC) was employed to evaluate the performance of our segmentation method. The average DSCs for the segmentation of the pituitary, left eye, right eye, left eye lens, right eye lens, left optic nerve, and right optic nerve were 90%, 94%, 93.5%, 84.5%, 84.3%, 80.3%, and 82.2%, respectively. CONCLUSION: We developed a new network-based approach for rapid and accurate CT image segmentation of the eyes and surrounding organs. This method is accurate and efficient, and is suitable for clinical use.

[The basic structure of heavy-ion tumor therapy facility].

Heavy-ions have the similar characteristic of depth-dose distribution with protons, but exhibit enhanced physical and radiobiological benefits. With increasing development in technical and clinical research, more facilities are being installed in the world. At the same time, many critical techniques of heavy-ion therapy facility were optimized and completed. This paper classified and reviewed the basic structure of heavy-ion system equipments, especially the accelerator, gantry, nozzle , TPS.

An adaptive window-setting scheme for segmentation of bladder tumor surface via MR cystography.

This paper proposes an adaptive window-setting scheme for noninvasive detection and segmentation of bladder tumor surface in T(1)-weighted magnetic resonance (MR) images. The inner border of the bladder wall is first covered by a group of ball-shaped detecting windows with different radii. By extracting the candidate tumor windows and excluding the false positive (FP) candidates, the entire bladder tumor surface is detected and segmented by the remaining windows. Different from previous bladder tumor detection methods that are mostly focusing on the existence of a tumor, this paper emphasizes segmenting the entire tumor surface in addition to detecting the presence of the tumor. The presented scheme was validated by ten clinical T(1)-weighted MR image datasets (five volunteers and five patients). The bladder tumor surfaces and the normal bladder wall inner borders in the ten datasets were covered by 223 and 10,491 windows, respectively. Such a large number of the detecting windows makes the validation statistically meaningful. In the FP reduction step, the best feature combination was obtained by using receiver operating characteristics or ROC analysis. The validation results demonstrated the potential of this presented scheme in segmenting the entire tumor surface with high sensitivity and low FP rate. This study inherits our previous results of automatic segmentation of the bladder wall and will be an important element in our MR-based virtual cystoscopy or MR cystography system.

A statistical study of the factors influencing the extent of respiratory motion blur in PET imaging.

Respiratory motion results in significant motion blur in thoracic and abdomen PET imaging. The extent of respiratory motion blur is mainly correlated with breathing amplitude, tumor size and location. In this paper we introduce a statistical study to quantitatively show the factors influencing the extent of respiratory motion blur in thoracic PET images. The study is centered on two regression models, one is linked with motion blur induced loss of mean intensity(LMI), tumor motion magnitude and tumor size, and another is to investigate the influence of tumor location, patient gender and patient height on tumor motion magnitude. We use the blur identification and image restoration technique to estimate the tumor motion and compute the LMI. The regression model was validated by simulation and phantom data before extended to 39 cases of clinical lung tumor PET images corrupted with blurring artifact. Results show that the motion magnitude of lung tumor during breathing is 10.9±3.7mm in transaxial plane, and it is significantly greater in lower lung lobes than in upper lobes. The LMI is 7.1±2.4% in the region of interest (ROI) above 40% of the image's maximum intensity. The least-square estimate of regression equations demonstrates that LMI is proportional to tumor motion magnitude and is inversely proportional to tumor size; the two factors play the same role in determining the extent of respiratory motion blur in thoraco-abdominal PET imaging. The location of tumor was shown as the major factor determining its motion magnitude, while the influencing of patient gender and height on tumor motion was not shown significant.

Low-dose CT reconstruction via edge-preserving total variation regularization.

High radiation dose in computed tomography (CT) scans increases the lifetime risk of cancer and has become a major clinical concern. Recently, iterative reconstruction algorithms with total variation (TV) regularization have been developed to reconstruct CT images from highly undersampled data acquired at low mAs levels in order to reduce the imaging dose. Nonetheless, the low-contrast structures tend to be smoothed out by the TV regularization, posing a great challenge for the TV method. To solve this problem, in this work we develop an iterative CT reconstruction algorithm with edge-preserving TV (EPTV) regularization to reconstruct CT images from highly undersampled data obtained at low mAs levels. The CT image is reconstructed by minimizing energy consisting of an EPTV norm and a data fidelity term posed by the x-ray projections. The EPTV term is proposed to preferentially perform smoothing only on the non-edge part of the image in order to better preserve the edges, which is realized by introducing a penalty weight to the original TV norm. During the reconstruction process, the pixels at the edges would be gradually identified and given low penalty weight. Our iterative algorithm is implemented on graphics processing unit to improve its speed. We test our reconstruction algorithm on a digital NURBS-based cardiac-troso phantom, a physical chest phantom and a Catphan phantom. Reconstruction results from a conventional filtered backprojection (FBP) algorithm and a TV regularization method without edge-preserving penalty are also presented for comparison purposes. The experimental results illustrate that both the TV-based algorithm and our EPTV algorithm outperform the conventional FBP algorithm in suppressing the streaking artifacts and image noise under a low-dose context. Our edge-preserving algorithm is superior to the TV-based algorithm in that it can preserve more information of low-contrast structures and therefore maintain acceptable spatial resolution.

Respiratory motion blur identification and reduction in ungated thoracic PET imaging.

Respiratory motion results in significant motion blur in thoracic positron emission tomography (PET) imaging. Existing approaches to correct the blurring artifact involve acquiring the images in gated mode and using complicated reconstruction algorithms. In this paper, we propose a post-reconstruction framework to estimate respiratory motion and reduce the motion blur of PET images acquired in ungated mode. Our method includes two steps: one is to use minmax directional derivative analysis and local auto-correlation analysis to identify the two parameters blur direction and blur extent, respectively, and another is to employ WRL, à trous wavelet-denoising modified Richardson-Lucy (RL) deconvolution, to reduce the motion blur based on identified parameters. The mobile phantom data were first used to test the method before it was applied to 32 cases of clinical lung tumor PET data. Results showed that the blur extent of phantom images in different directions was accurately identified, and WRL can remove the majority of motion blur within ten iterations. The blur extent of clinical images was estimated to be 12.1 ± 3.7 mm in the direction of 74 ± 3° relative to the image horizontal axis. The quality of clinical images was significantly improved, both from visual inspection and quantitative evaluation after deconvolution. It was demonstrated that WRL outperforms RL and a Wiener filter in reducing the motion blur with one to two more iterations. The proposed method is easy to implement and thus could be a useful tool to reduce the effect of respiration in ungated thoracic PET imaging.

Volume-based features for detection of bladder wall abnormal regions via MR cystography.

This paper proposes a framework for detecting the suspected abnormal region of the bladder wall via magnetic resonance (MR) cystography. Volume-based features are used. First, the bladder wall is divided into several layers, based on which a path from each voxel on the inner border to the outer border is found. By using the path length to measure the wall thickness and a bent rate (BR) term to measure the geometry property of the voxels on the inner border, the seed voxels representing the abnormalities on the inner border are determined. Then, by tracing the path from each seed, a weighted BR term is constructed to determine the suspected voxels, which are on the path and inside the bladder wall. All the suspected voxels are grouped together for the abnormal region. This work is significantly different from most of the previous computer-aided bladder tumor detection reports on two aspects. First of all, the T (1)-weighted MR images are used which give better image contrast and texture information for the bladder wall, comparing with the computed tomography images. Second, while most previous reports detected the abnormalities and indicated them on the reconstructed 3-D bladder model by surface rendering, we further determine the possible region of the abnormality inside the bladder wall. This study aims at a noninvasive procedure for bladder tumor detection and abnormal region delineation, which has the potential for further clinical analysis such as the invasion depth of the tumor and virtual cystoscopy diagnosis. Five datasets including two patients and three volunteers were used to test the presented method, all the tumors were detected by the method, and the overlap rates of the regions delineated by the computer against the experts were measured. The results demonstrated the potential of the method for detecting bladder wall abnormal regions via MR cystography.

4D CT image simulation using B-Spline deformable model and cosine interpolation of deformation field.

Four-dimensional computed tomography(4D CT) is significant in radiotherapy treatment planning for thorax and upper abdomen to take their motion induced by respiration into consideration, but its high radiation dose becomes a major concern and impedes its wide application. To solve the problem, we propose an image interpolation approach to get 4D CT simulation images. We simulate 4D CT images at arbitrary intermediate phases by B-Spline deformable model with cosine interpolation of the deformation field, which is obtained by deformable registration of two CT images at end-exhale and end-inhale phases. The mean of absolute differences computed between actual 4D CT images and simulation ones is used to evaluate the accuracy of simulation. Our experiment results show that both linear interpolation and cosine interpolation with proper parameters perform well and the latter performs a little better than the former in general.

Sparse p-norm Nonnegative Matrix Factorization for clustering gene expression data.

Nonnegative Matrix Factorization (NMF) is a powerful tool for gene expression data analysis as it reduces thousands of genes to a few compact metagenes, especially in clustering gene expression samples for cancer class discovery. Enhancing sparseness of the factorisation can find only a few dominantly coexpressed metagenes and improve the clustering effectiveness. Sparse p-norm (p > 1) Nonnegative Matrix Factorization (Sp-NMF) is a more sparse representation method using high order norm to normalise the decomposed components. In this paper, we investigate the benefit of high order normalisation for clustering cancer-related gene expression samples. Experimental results demonstrate that Sp-NMF leads to robust and effective clustering in both automatically determining the cluster number, and achieving high accuracy.

On alpha-divergence based nonnegative matrix factorization for clustering cancer gene expression data.

OBJECTIVE: Nonnegative matrix factorization (NMF) has been proven to be a powerful clustering method. Recently Cichocki and coauthors have proposed a family of new algorithms based on the alpha-divergence for NMF. However, it is an open problem to choose an optimal alpha. METHODS AND MATERIALS: In this paper, we tested such NMF variant with different alpha values on clustering cancer gene expression data for optimal alpha selection experimentally with 11 datasets. RESULTS AND CONCLUSION: Our experimental results show that alpha=1 and 2 are two special optimal cases for real applications.

Reducing microarray data via nonnegative matrix factorization for visualization and clustering analysis.

In microarray data analysis, each gene expression sample has thousands of genes and reducing such high dimensionality is useful for both visualization and further clustering of samples. Traditional principal component analysis (PCA) is a commonly used method which has problems. Nonnegative Matrix Factorization (NMF) is a new dimension reduction method. In this paper we compare NMF and PCA for dimension reduction. The reduced data is used for visualization, and clustering analysis via k-means on 11 real gene expression datasets. Before the clustering analysis, we apply NMF and PCA for reduction in visualization. The results on one leukemia dataset show that NMF can discover natural clusters and clearly detect one mislabeled sample while PCA cannot. For clustering analysis via k-means, NMF most typically outperforms PCA. Our results demonstrate the superiority of NMF over PCA in reducing microarray data.