This PDF file contains the front matter associated with SPIE Proceedings Volume 11792, including the Title Page, Copyright information, and Table of Contents.

Computer-aided diagnosis (CAD) has gained considerable attention for breast cancer screening owing to its high diagnostic efficiency and satisfactory accuracy. However, it has been revealed that traditional CAD systems for mammography are vulnerable to dense breast tissue, which could hide underlying tumors. To resolve this issue, we devised a learning scheme that equips the U-Net backbone with a well-designed attention mechanism to suppress the over-detection rate for nongland mammary regions in dense breast tissue and applied to the CAD for breast ultrasound (BUS) images. The proposed method has two stages: initial mammary gland segmentation, which involves the selection of a region in the mammary gland where a tumor may occur; then tumor region segmentation, wherein the attention U-Net detects tumor regions by characterizing the selected mammary gland probability map as a spatial attention map, drawing selective attention to mammary gland tissues. We evaluated the proposed tumor detection scheme on several public BUS image datasets. Comparative results demonstrate that the proposed approach achieves the best performance in most conditions. Notably, when considering the percentage of all actual tumors that were correctly segmented, the proposed method showed a tumorwise accuracy performance of 92.7%.

We proposed a method that detects DLBCL (Diffuse Large B-Cell Lymphoma) regions from a H&E stained whole slide pathology image by measuring the size of each nucleus. It is known that DLBCL cells would have about 2 to 3 times larger nuclei than typical lymphocytes. One can hence detect DLBCL regions by detecting every cell nucleus in a given H&E stained pathology image and describing the spatial distribution of the large nuclei. For the detection of cell nuclei, we employ a U-Net and a Bayesian U-Net. We describe the details of the proposed method and report the experimental results, which demonstrate the proposed method works well in the DLBCL regions.

Recently, the need for liver transplantation has increased as the number of liver cancer and liver cirrhosis patients increases. The preoperative measurement of the liver volume of the donor is very important. The liver volume is one of analysis factors to predict liver function. However, the current process of liver volume is manually measured by radiologist from CT data, and it takes a lot of time and effort. In this paper, we propose a Deep 3D Attention U-Net for the whole liver segmentation that learns to focus on liver structures of varying shapes and sizes. In addition, the whole liver volume was calculated in voxel units using the segmentation result. The liver segmentation studies of the 266 patients are randomly assigned into train, validation and test sets, with a split ratio of 80%, 10% and 10% of total amount of patients, respectively. The results of liver segmentation achieved sensitivity of 0.914, the specificity of 0.999, and the dice similarity coefficient of 0.936. The relationship analysis of the liver volume showed the correlation coefficient r of 0.853 between manually measured liver volume and calculated liver volume using segmentation result. The results of liver volume measurements through whole liver segmentation based on Deep 3D Attention U-Net were similar to a reliable level.

Autoimmune diseases are an abnormal immune response of human body, which could cause mild symptoms like low grade fever, or severe reactions including damage to joints and muscles or even causing cancer. The early symptoms of autoimmune diseases are generally the same as common illnesses and occasionally occurring, so the physician will apply various tests to determine the existence of autoimmune diseases. Among the commonly used test methods, antinuclear autoantibody (ANA) screening is the one most often used. However, the ANA screening interpretation is a laborintensive process, requiring hours of physician per day to read the specimens of ANA testing. Since neural networks granted new progress in 2012, deep convolutional networks in medical applications have caught physicians' interests. Current approaches of automatic process to autoimmune diseases testing are focused at single class annotation dataset, such as I3A and ICPR 2012 dataset. This study proposes a method using mask R-CNN in mixed pattern dataset to segment the cells in instance level, and classify cells into different cell cycles for further classification research. The proposed method achieves 89% segmentation accuracy and 95.07% cell cycle classification accuracy.

Diffuse lung diseases (DLD) are widely distributed in lungs. Because the opacity patterns of DLD on CT images are complex and various, the diagnostic results may be different between doctors depending on their experience and subjective decision on them. In order to solve this problem, performing image analysis using CAD (Computer-Aided Diagnosis) systems attracts attention. To achieve high performance in diagnosis by using these CAD systems, it is necessary to first perform lung region extraction as preprocessing for limiting the target domain. However, by using the existing systems, it is difficult to extract lung regions from all five typical shadow patterns of DLD and normal lungs. In this study, we aimed to extract lung regions from CT slices containing DLD shadows using the U-net for improving the CAD performance.

Supervised learning for image segmentation requires annotated images. However, image annotation has the problem that it is time-consuming. This problem is particularly significant in the erector spinae muscle segmentation due to the large size of the muscle. Therefore, this study considers the relationship between the number of annotated images used for training and segmentation accuracy of the erector spinae muscle in torso CT images. We use Bayesian U-Net, which has shown high accuracy in thigh muscle segmentation, for the segmentation of the erector spinae muscle. In the network training, we limit the number of slices for each case and the number of cases to 100%, 50%, 25%, and 10%. In the experiment, we use 30 torso CT images, including 6 cases for the test dataset. Experimental results are evaluated by the mean Dice value of the test dataset. Using 100% of the slices per case, the segmentation accuracy with 100%, 50%, 25%, and 10% of the cases was 0.934, 0.927, 0.926, and 0.890, respectively. On the other hand, using 100% of the cases, the segmentation accuracy with 100%, 50%, 25%, and 10% of the slices per case was 0.934, 0.934, 0.933, and 0.931, respectively. Furthermore, the segmentation accuracy with 100% of the cases and 10% of the slices per case was higher than that of the previous method. We showed that it is feasible to achieve high segmentation accuracy with a limited number of annotated images by selecting several slices from a limited number of cases for training.

Recently, it has been shown that the uncertainty can be estimated by integrating Monte Carlo (MC) dropout [1] in the neural network structure in deep learning. Our group proposed Bayesian U-Net [2], which integrates MC dropout with U-Net [3] architecture, and investigated the relationship between the segmentation accuracy (Dice) and the uncertainty inferred by MC dropout. High correlation between the uncertainty and the errors in the automatic segmentation of musculoskeletal structures in CT images was obtained. In a different study, the integration of the segmentations obtained by voting based on prediction probabilities by convolutional neural networks (CNNs) trained for each anatomical plane independently [4] could further improve the segmentation accuracy and reduced the calculation time. However, as far as we know, the uncertainty-based integration has not been investigated yet. In this paper, we applied 3 Bayesian U-Nets, each trained on 2D slices in one of the three anatomical planes, to two-phase contrast-enhanced CT images of 48 cases (97 images) with manual segmentation of 17 organs. We report the variations in the uncertainties with respect to the anatomical planes, and finally evaluate the multiplanar integration-based predictions compared with single plane-based predictions. The segmentation accuracy, represented by Dice coefficient (DC), was significantly improved by the uncertainty-based integration in 9 organs (p<0.05) compared with segmentations obtained only from the axial plane. All segmentation results showed a negative correlation between the uncertainty and DC. We found that the segmentation accuracy could be improved by integrating the multiplanar segmentation results based on uncertainty estimation.

As one of image pre-processing method to detect, recognize, and estimate lesion or characteristic region in medical image processing, there are many studies improved performance and precision of processing by contrast enhancement or super-resolution. However, it is not clarified how condition is better to apply these methods. Therefore, we experimented and discussed on affect for color laparoscopic image quality by the difference of contrast enhancement method. As a result, we obtained knowledge of high similarity among patterns of adaptive histogram equalization in three methods. However, under these conditions, in the case of considering the region segmentation, it is not clarified how processing precision is better. In this paper, first we processed the contrast enhancement for the color laparoscopic frame image cut from surgery video under laparoscopy. Next, we processed super-resolution for generated image. Finally, we compared and discussed by Peak Signal to Noise Ratio (PSNR), Structural SIMilarity (SSIM), and texture features for contrast.

We analyze fixation and saccade using an extension of a convolutional neural network (CNN) model and compare the results to a conventional modeling method (gazeNet). Unlike the conventional method, ours can easily be applied to realtime analysis of eye movements during data acquisition since it can feed the results back to trainees for image interpretation in real time. Eye movement data was divided into “fixation” and “saccade” sections using the open data sets, “Lund2013” and “GazeCom,” that are available via the internet and which can be used as validation data for interpreting eye movements while viewing medical images. Images to be input into our Deep CNN model, DCNN, were created by drawing path lines from 12 consecutive gaze points over a period of 0.1 s assuming 120 Hz measurements with appropriate downsampling of the data (500 Hz for Lund2013 and 250 Hz for GazeCom). Our DCNN model was shown to be largely superior to gazeNet, yielding high sensitivity (97.7% for Lund2013 and 98.2% for GazeCom) and specificity (86.4% and 93.8%, respectively). Our findings show that eye movement data classification was generally more accurate for our DCNN model than for the previously reported model.

In this study, we aimed to classify lung cancers in chest CT images into adenocarcinoma (AD) and squamous cell carcinoma (SQ) using 3D Convolutional Neural Networks (CNN), and to visualize grounds used in the classification process by CNN. Although CNN is a powerful tool for classifying types of lung cancers, it does not provide grounds for decision explicitly, and there is a possibility that doctors and patients may not be satisfied with the decision by CNN. First, we developed a CNN based classifier to classify lung tumors into AD and SQ. The recognition rate of the proposed method was 69.9 ± 3.8%. Furthermore, the grounds of the classification by CNN was visualized by using Gradient-weighted Class Activation Mapping (Grad-CAM)[1].

Acute kidney injury (AKI) is associated with increased morbidity and mortality in intensive care units (ICU). The sudden episode of kidney failure may lead to end-stage renal disease (ESRD) or deaths, and has been related to significantly increasing costs of ICU admissions and treatments. Early prediction of AKI inpatient mortality will help decision-making, and benefit resource allocation in ICU. Therefore, it is crucial to develop an early warning system for AKI prediction. We aimed to create a more comprehensive predictive model for 1-year AKI mortality. A cohort of 2,247 patients with AKI was assembled, of which the in-hospital mortality was 36.67%. Longitudinal data of each patient were collected from the Medical Information Mart for Intensive Care III (MIMIC-III) dataset. An interpretable XGBoost risk model was developed and validated by 10-fold cross validation. Model predictors included 11 routinely collected AKI-related laboratory measurements, 8 complications of AKI, and demographic data. An artificial neural network (ANN) model was also developed in parallel for comparison. The XGBoost model demonstrated an area under the receiver-operating characteristic curve (AUC) of 0.83, which was superior to ANN (AUC = 0.79). Our model was able to predict mortality of AKI in ICU with high accuracy. Our model can predict 1-year AKI mortality. Furthermore, it had great potential for identifying at-risk patients in ICU. These findings indicated that our approach might offer opportunities for better resource utilization and better administration of AKI.

In current society, height is one of the conditions for many people to measure the pursuit of perfection. The focus is no more than judging whether the growth plate of the bone is healed or not, and whether it will grow. Generally, the growth plate will heal with age, and the bone will stop growing after the healing is complete. Presently, in clinical bone age analysis, the most famous method is still GP method, which published in 1959 by Greulich and Pyle et al. They used normal left palm and wrist X-ray images to be the references in different ages, and discriminated the difference bone ages between normal person and examinee. However, manual interpretation is a boring task and time-consuming. The faster and accurate automatic estimation for bone age analysis is necessary. This study is based on deep neural network (DNN) algorithm. The Python programming modules, InceptionResNetV2 and Xception, are respectively used to implement ours proposed computer-aided system of bone age estimation. We also apply into the threshold segmentation and major axis correction method to assist the DNN training procedure, which can effectively remove redundant noise around the hand bone in X-ray images. In the experiments, there are 12,611 X-ray images in our database. During threshold segmentation, there are only 14 segmentation fault cases, accounting for 0.1% of total cases. Furthermore, the proposed system with DNN module can obtain a high accuracy rate and a small loss function in the training set. The proposed system in this study effectively enhances the bone age estimation. In the future, different DNN modules can be tried to improve the performance of ours system.

Alzheimer’s disease (AD) is a progressive brain disease that causes a different pattern of brain atrophy from normal aging. Early identification of AD is crucial since the progression of the disease can be slowed down by medication. In the field of image recognition, its accuracy has been significantly improved by using convolutional neural networks (CNNs). Similarly, in the field of medical image processing, researches on the diagnostic support using CNN have been studied. In this paper, we propose an AD classification method using CNN, inspired by the success of CNNs in brain age estimation. Through experiments using a large-scale database, we demonstrate the effectiveness of our proposed method.

Knee injury emerges as one of the most common diseases, causing dislocation of knee joints, immobility, etc., in which the anterior cruciate ligament (ACL) injury is the most common one. The development of various artificial intelligence (AI) frameworks gained enormous attention in many areas, including injury prediction and health management via medical image analysis. The objective of the current study is to focus on a comprehensive high accurate prediction of ACL injury based on MRI medical images, and also demonstrate the ability of AI in practical and outline conceptual prediction and diagnosis frameworks for other types of knee injuries in the future. Our dataset comprised of knee MRI reports from Cho Ray Hospital, Vietnam which are composed of ACL and non-ACL injury patients. The MRI images were used as supporting data in the deep learning classification model with DenseNet-121 algorithm. The successful establishment of an ACL injury diagnosis model from MRI will pave the way for us to develop more diagnostic models of other injuries in the body as well as the prediction of bone diseases.

Convolutional neural networks (CNNs) have shown significant success in image recognition and segmentation. Based on a CNN-like U-Net architecture, such a model can effectively predict subcellular structures from transmitted light (TL) images after learning the relationships between TL images and fluorescent-labeled images. In this paper, we focused on building corresponding models of subcellular mitochondrial structures using the CNN method and compared the prediction results derived from confocal microscopic, Airyscan microscopic, z-stack, and time-series images. With multi-model combined prediction, it is possible to generate integrated images using only TL inputs, which reduces the time required for sample preparation and increases the temporal resolution. This enables visualization, measurement, and understanding of the morphology and dynamics of mitochondria and mitochondrial DNA.

We present a novel approach to movability assessment on physiotherapy for shoulder periarthritis via fine-grained 3D Residual Networks (R3D) deep learning. The unique deep neural networks is able to automatically extract the spatiotemporal features from the RGB-D videos. In our preliminary studies, we have a set of VR sports games customized for the immersive and interactive sports environment, to regulate the patient’s rehabilitation exercises. In this way, acquisition of RGB-D action videos can be more specific to the subject and defined movements; and fine-grained feature discrimination of the same subject can be better achieved from the longitudinal study, to increase the accuracy of therapeutic assessment.

The purpose of this study is to use a deep learning model to identify the possibility of lesions in the cervix and to evaluate the efficient image preprocessing in order to diagnose diverse types of cervix in form. The study used 4,107 normal photographs of uterine cervix and 6,285 abnormal photographs of uterine cervix. Under the same size condition, to see if which method is more effective to performance either removal of the vaginal wall area or diagnosing cervical cancer including the vaginal wall area, two types of image preprocessing were resized to square. The average accuracy of cropped cases is 94.15%. The average accuracy of the filled cases is 93.41%.

Artificial intelligence (AI) training courses often require prerequisites such as calculus or statistics. It is hence challenging to design and develop an introductory AI course for students of secondary education. This research intends to develop a medical AI course, provide high school students with an overview of deep learning applications in medical image analysis, and inspire them to pursue careers in the field of medical AI. We designed a 20-hour course, including lectures and two hands-on projects based on medical image classification. The proposed courses provided medical AI disciplines and built up their knowledge from basic to advanced levels. During the ten-day online courses, all the students were fully engaged and gave us positive feedback. The students endeavored to complete the experimental study in training, testing, and hypothesis of medical images application in the course. Their performance exceeded all expectations, for they did further analysis by tuning different hyperparameters. We designed a course evaluation form, which suggested that the students found it essential and expected to interact with the instructors. The results indicate that combining lectures with hands-on sessions would lead to evidently better achievement in terms of high school students’ medical AI knowledge and positive attitudes while addressing real-world problems in the projects. Through this innovative education model, high school students regained their enthusiasm and were encouraged to cultivate their medical AI skills through self-learning while finishing the project. We conclude that this course could be successfully applied to interdisciplinary education in high school.

Respiratory auscultation can help healthcare professionals detect abnormal respiratory conditions if adventitious lung sounds are heard. The state-of-the-art artificial intelligence technologies based on deep learning show great potential in the development of automated respiratory sound analysis. To train a deep learning-based model, a huge number of accurate labels of normal breath sounds and adventitious sounds are needed. In this paper, we demonstrate the work of developing a respiratory sound labeling software to help annotators identify and label the inhalation, exhalation, and adventitious respiratory sound more accurately and quickly. Our labeling software integrates six features from MATLAB Audio Labeler, and one commercial audio editor, RX7. As of October, 2019, we have labeled 9,765 15- second-long audio files of breathing lung sounds, and accrued 34,095 inhalation labels,18,349 exhalation labels, 13,883 continuous adventitious sounds (CASs) labels and 15,606 discontinuous adventitious sounds (DASs) labels, which are significantly larger than previously published studies. The trained convolutional recurrent neural networks based on these labels showed good performance with F1-scores of 86.0% on inhalation event detection, 51.6% on CASs event detection and 71.4% on DASs event detection. In conclusion, our results show that our proposed respiratory sound labeling software could easily pre-define a label, perform one-click labeling, and overall facilitate the process of accurately labeling. This software helps develop deep learning-based models that require a huge amount of labeled acoustic data.

Laparoscopic surgery provides for patients such advantages as a small incision range and quick postoperative recovery. Unfortunately, surgeons struggle to grasp 3D spatial relationships in the abdominal cavity. Methods have been proposed to present the 3D information of the abdominal cavity using AR or VR. Although 3D geometrical information is crucial to perform such methods, it is difficult to reconstruct dense 3D organ shapes using a feature-point-based 3D reconstruction method such as structure from motion (SfM) due to the appearance characteristics of organs (e.g., texture-less and glossy). Our research solves this problem by estimating depth information from laparoscopic images using deep learning. We constructed a training dataset from both RGB and depth images with an RGB-D camera, implemented a depth image generator by applying a generative adversarial network (GAN), and generated a depth image from a single-shot RGB image. By calibration with a laparoscopic camera and an RGB-D camera, the laparoscopic image was transformed to an RGB image. We generated depth images by inputting the transformed laparoscopic images into a GAN generator. The scale parameter of the depth image with real-world dimensions was calculated by comparing the depth value and the 3D information estimated by SfM. Consequently, the density of the organ model increased by back-projecting the depth image to the 3D space.

Recent development of compressed sensing (CS) and deep learning (DL) brought a significant progress in image reconstruction for sparse-view CT and low-dose CT. However, there still exist a strong demand in further improving image quality. We propose a new framework for image reconstruction in sparse-view CT and low-dose CT, which significantly outperforms CS and DL in terms of image quality. This advantage originates from combining CS and DL in a successful way as described below, thereby leading to compensating for each other’s weakness. The proposed framework is based on the following principle. First, CS image reconstruction using TV (or Nonlocal TV) regularization is performed with prespecified M different values of regularization parameters (β₁, β₂, ---,β_M), which generates M reconstructed images (z₁,z₂, - --,z_M) with varying degree of TV smoothing. Next, the TV images (z₁,z₂, ---,z_M) together with a FBP reconstruction (no smoothing) y are inputted into CNN having M+1 input channels and single output channel. The final reconstructed image is obtained as the output of CNN. With respect to the learning of network, CNN parameters (weights and biases) are estimated by minimizing an MSE loss function using learning data, i.e. a set of M+1 input images and corresponding answer image. In our previous work [11], we have already proposed a similar framework for the case where the number of input TV image is one. However, we expect that increasing the number of input images as mentioned above will further improve image quality. In this work, we have investigated such a new extension. Intuitively, the proposed method is based on combining good parts in M+1 input images to synthesizing a higher quality image, and this synthesis is performed by using DL. We have performed a simulation study using a dataset of clinical abdominal CT images for 2-D low-dose CT and 2-D sparse-view CT. The result demonstrates that the proposed combined approach is able to significantly improve image quality compared to the case where CS or DL was used alone, both in terms of numerical evaluation (RMSE and SSIM) and visual evaluation.

In order to improve the accuracy of surgical analysis and clinical diagnosis, it is very important to obtain brain structure images. And magnetic resonance imaging (MRI) is one of the most commonly used methods. Moreover, High-Resolution (HR) MR image with smaller pixel size provides more important structure and texture details [1] and helps early diagnosis and subsequent analysis. But in fact, it is relatively difficult to obtain high-quality MR images due to many factors, such as hardware equipment, imaging time, required signal to noise ratio (Signal to Noise Ratio, SNR) and motion artifacts, etc. Generally speaking, brain MR images are often obtained with thick slice thickness and lower image quality to reduce the scanning cost and sampling time. However, this is not conducive to further medical analysis. For decades, Super- Resolution (SR) related technologies have been used to improve the quality of MR images to restore important structural information and facilitate clinical diagnosis. Also, as deep learning (DL) develops significantly nowadays, DL-based methods are widely applied to SR issues. Therefore, we propose a convolution neural network (CNN) SR method called Channel Splitting Edge-guided Residual Network (CSERN). Besides, we combine our method with a novel accelerating imaging method called Single-frequency Excitation WideBand (SE-WB) MRI and design a loss function to achieve higher performance on several index such as structural similarity index (SSIM) and Peak signal-to-noise ratio (PSNR).

We propose a probe localization method only from ultrasound (US) image sequences using deep learning for three-dimensional (3D) US image reconstruction. The proposed method employs a convolutional neural network (CNN) to estimate the motion of the probe from two US images. Our CNN architecture consists of two parts: inplane and out-of-plane probe motion estimation. Two loss functions are introduced to guarantee the consistency of estimated motion of the probe between multiple frames. Through experiments, we demonstrate that the proposed method exhibits efficient performance on probe localization compared with the conventional method.

Compressed sensing (CS) image reconstruction in CT suffers from the drawbacks such as 1) appearance of staircase artifacts and 2) loss in image textures and smooth intensity changes. These drawbacks stem from the fact that CS is based on approximating the image by a piecewise-constant function. To overcome this drawback, we have already proposed a framework to improve image quality in CS using deep learning. In this framework, FBP reconstructed image and CS (TV or Nonlocal TV) reconstructed image are inputted to CNN with two input channels and single output channel, and a final reconstructed image is obtained by the output of CNN. Parameters (weight and bias) of CNN together with a regularization parameter of CS are estimated by minimizing an average least-squares loss function by using learning data, i.e. a set of triplet of degraded FBP reconstruction, CS reconstruction, and answer image. In this paper, this framework is extended to 3-D image reconstruction in helical cone-beam CT operated with lowdose scanning protocol. Parameters (weight and bias) of CNN together with a regularization parameter of CS are estimated by minimizing an average least-squares loss function by using learning data, i.e. a set of triplet of degraded FBP reconstruction, CS reconstruction, and answer image. In this paper, this framework was extended to 3-D image reconstruction in helical cone-beam CT operated with lowdose scanning protocol. The extension was done in the following way. First, we prepare N different 2-D denoising CNN (CNN₁, CNN₂, . . . , CNNN ) dependent on the slice position n. Each slice of the short-scan FDK reconstruction without denoising y_i and with 3-D TV (or Nonlocal TV) denoising z_i are inputted to CNNn with the closest slice index n, which yields a corresponding output image for each slice x_i . The final reconstructed image is obtained by stacking every slice x_i (i = 1, 2, . . . , I).

This paper proposes an unpaired medical image translation framework between portal-venous phase and non-contrast CT volumes. Image-to-image translation has immense potential application values in medical image analysis fields, such as segmentation. Currently, many deep learning-based segmentation methods have been proposed on contrast-enhanced CT volumes. However, for the patients who have contrast medium allergy, only non-contrast CT is available. Thus, segmentation using non-contrast CT volumes is also an important task. Image translation from non-contrast CT to contrast-enhanced CT is an alternative to solve this problem. In this work, we employed the cycle-consistent adversarial network (CycleGAN) and unpaired image-to-image network (UNIT) for image translation. To evaluate the translation performance for multi-organ segmentation, we trained a segmentation model using contrast-enhanced CT images with U-Net. Our experimental results show that image translation has a positive influence on multi-organ segmentation. The segmentation actuaries greatly improved by applying the image translation.

We propose a method to transfer a given pathology image stained by some immunostaining to a H&E stained one. When one construct a classifier that estimates the subtype of malignant lymphoma from a given H&E stained pathology image, one needs a set of training H&E stained whole slide images in which the tumor regions are annotated. The annotation is not easy and requires large human resources. Here, it is known that some immunostaining stains only some specific tumor cells and the tumor region detection from the immunostained images is straightforward. It means once you transfer the immunostained images to H&E stained ones, you can easily obtain a set of virtually H&E stained images with annotation of tumor regions. In this manuscript, we report on the proposed method and experimental results of stain transfer from CD20 stained images to H&E stained ones.

Two key research issues are addressed: (i) A semantic representation and interpretation framework by using a lightweight self-supervised learning approach, namely the Context-Free Grammar and Push-Down Automaton; and (ii) A mobile phone App implementation of B-mode medical ultrasound imaging with a handheld probe, which can make use of the learned semantic features of scanned images for future home-based health screening.

The sinogram inpainting based methods such as normalized metal artifact reduction (NMAR) shows good performance in reducing the metal artifacts. However, these methods wipe away the structures near the metal which are severely corrupted by artifacts. The proposed method using a weight to lower metal corrupted pixels during the iteration instead of using an inpainting method to conserve the structures. Then, the proposed method was complementarily used with NMAR to generate a corrected image using the frequency split method. The titanium inserted XCAT phantoms were simulated from the 80kVp of energy. The metals were segmented from the filtered backprojection image using the threshold from the projection data. Afterwards, they were projected to acquire the metal sinogram index and the metal subtracted projection data. A weight map is generated from the scale factor and the log of the original raw data which has the metal index. The metal subtracted projection data were iteratively reconstructed. During the iteration, the step size of the update term was affected by a weight map, and the contributions of the metal present sinogram were lowered. Finally, using the frequency split approach, the corrected image was generated from the proposed and NMAR results. It was possible to lower the artifacts while conserving the structures in the proposed method. This work indicates the weighted iterative reconstruction can be complementarily used with the existing FSMAR approach.

Freehand 3D ultrasound imaging using a 1D transducer array has been widely investigated. Speckle decorrelation-based elevational displacement estimation is often applied. Generally, the correlation coefficient (C.C.) of two regions of interest is mapped to the beam pattern which can be utilized to estimate the elevational displacement. However, performance has been limited due to several factors, including the inherent variance of pure speckle patterns. In this study, we propose a more robust and accurate approach that utilizes a speckle generating ultrasound gel pad, singular value decomposition (SVD), and machine learning for improving estimating performance. First, a 0.5-cm-thick speckle generating gel pad was used to produce homogeneous patterns with statistically fully developed scatterers. Second, calculations of the decorrelation curves were improved with the introduction of SVD method. Third, the two-layer artificial neural networks were utilized for estimation. With training by totally 4600 motion data with frame space of 0.01 mm and 0.1° respectively, our estimator achieves 0.906 precision while estimating the motion type, as well as the average error of displacement / rotation movement is 0.0002 mm and 0.004° respectively.

The volume exploration refers to the observation of internal structures by manipulating visualization widgets such as a slicing plane. Since the advent of a commercial holographic display, the demand for 3D contents is also increasing in the medical visualization area. Holographic displays deliver a set of viewpoints through the lenticular lens so that a user can observe different sides of the object without wearing a pair of special glasses. Moreover, it comes natural for the user to perform the volume exploration using mid-air interaction. In this work, we present a novel volume exploration technique for holographic display. To accelerate the hologram rendering, we adopt the volume rendering pipeline to the holographic display mechanism. Also, to support a spatial interaction for volume exploration, we develop a cut-and-annotate interaction between volume object and point cloud. The proposed voxel classification algorithm produces a set of class properties for delineating different structures of the 3D object. We demonstrate our volume exploration method on several volume data sets under real-time user interaction.

The goal of this study is to increase the measurement accuracy of the bladder volume for point-of-care ultrasound (POCUS). An algorithm that can utilize spatial information from inertial measurement units (IMUs) embedded in POCUS and ultrasound images to estimate bladder volume has been developed. So far, ultrasound scanning is a noninvasive technique for treatment and diagnosis in the hospital. Bladder volume determination in post-void residual (PVR) through ultrasound can help clinicians. However, the ultrasound machines with the ability of calculating volumes precisely in hospital are bigger and expensive than POCUS. The goal of this study is to improve the accuracy of bladder volume without expensive instruments. We use an on-the-shelf wireless hand-held convex probe (LU700C, LELTEK Inc, Taiwan) to collect bladder images. LU700C is also capable of providing real-time posture information to detect User behavior. To further enhance the accuracy, an extra IMU has been attached on scanner for collecting posture data conveniently at scanning. The original prolate ellipsoid formula-based algorithm calculates bladder volume with virtual caliper. The bladder phantoms are made by ourselves to further verify the accuracy. Each of the measurement of bladders were repeat three times to follow the accepted procedural. The results show integrate hand’s posture information with timestamp into sonogram frames during bladder scanning can improve accuracy of volume estimation effectively. The proposed algorithm implements in current devices using in bladders measurement performs significantly better than the existing ones. Our goals of this research are to improve the quality of clinical through software-update without any change of hardware and to bring sustainable healthcare for areas lacking of medical resources.

Cirrhosis of liver may cause structural distortion of entire liver by fibrosis and parenchymal nodules, in which the image findings in MR/CT images can be interpreted by shape, texture, volume, elasticity analysis and so on. To ease the workload of radiologists from interpretation of the numerous medical images, an online Computer-aided diagnosis (CAD) system on liver has been developing for quantifying the diagnosis of fibrosis. Several technologies regarding image processing and pattern recognition are introduced in this paper, including Shape, Texture, Volume, Elasticity Analysis and Edge computing with FPGA acceleration. The results by the novel methods on the non-invasive assessment of liver fibrosis as well as the role of our CAD scheme to replace traditional liver biopsy will be presented and discussed.

Diffusional kurtosis imaging (DKI) parameters were conventionally inferred by using signal model fitting. Recent researches by using machine learning showed improvement in parameter inference robustness. Especially the approach of training only with synthetic data, called synthetic Q-space learning (synQSL), has been proved to be have several advantages. In our previous study on synQSL, single level of Rician noise was used for simulating realistic signals in training data, but noise level of clinical data is supposed to be in mixture distribution. In this study, we investigated basic characteristics and usefulness of the mixture distribution noise in training data through experiments with synthetic and real data. Five-layered multi-layer perceptron (MLP) was used for synQSL. The MLP inputs were three values of logarithm of normalized signal decays, that is, we assume The output is single for diffusion coefficient 𝐷 or diffusional kurtosis 𝐾. Two different levels of Rician noise were simulated and were mixed to training dataset. Various levels and mixture ratios of synthetic data for training were examined by using synthetic test data and clinical data to obtain the basic characteristics. Synthetic data experiments showed that noise level matching yields robust inference for 𝐷 and 𝐾 also in mixed distribution of noise level, similar to single noise level training. Clinical data experiments showed that training with mixed noise level yielded less outliers (out-of-range values: 𝐾 < 0 or 𝐾 > 2) than single noise level training. Thus, the synQSL with mixture distribution noise has potential to improve robustness for DKI parameter inference in clinical data.

This paper deals with the so-called Metal Artifact Reduction (MAR) in CT. This problem aims at reconstructing a CT image with reduced metal induced artifact when the object contains metallic parts inside. We propose a new iterative reconstruction method to the MAR problem, which uses the L₁ norm for data fidelity term and Nonlocal TV regularization. In ordinary iterative reconstruction for CT, the least-squares error || A ^→x - ^→b|| ²₂ Is used as data fidelity term for image reconstruction. However, it is well-known that the least-squares criterion is sensitive to the existence of abnormal (inconsistent) data in the measurement ^→b, such as projection data passing through the metallic parts in this work. A simple reasonable method to identify the location of metallic parts in the sinogram and exclude the corresponding projection data from the data fitting is to use the L₁ norm error || A ^→x - ^→b|| ¹₁ . Furthermore, the power of proposed method to reduce the metal artifact can be significantly improved by adding Nonlocal Total Variation (NLTV) regularization term into the cost function. Compared to existing approaches to the MAR problem, the proposed method possesses the following attractive feature. Almost every approach to MAR consists of two-step computations. The first step detects the metallic parts in the sinogram and the second step performs image reconstruction after interpolating or excluding the projection data corresponding to the identified metallic parts. On the other hand, the proposed method consists of only a single computational step, i.e. single iterative minimization of a convex cost function, leading to smartly unifying the two steps into a single step.