Abstracts Track 2021

Nr: 36
Title:

SportBand: An Innovative Tool for Monitoring and Supporting the Physical Activity
Authors:

Dorota Dulko, Przemysław Łukasiewicz, Maciej Wierzbowski, Małgorzata Mielnik, Maciej Danielewicz, Rafał Banasiuk and Krzysztof Rykaczewski
Abstract: Human sweat is a biofluid which is very attractive for point-of-care health monitoring. State of the art sensors in the form of patches or other wearable technologies allow for the detection of sweat constituents. Moreover, there are a lot of options for monitoring everyday activities with recently made widely available fitness bands, smartwatches etc. Our solution comprises of those two solutions in one biodevice. SportBand is a wireless diagnostic tool to be used during physical activity. The objective of this device is the real-time monitoring of factors like heart rate, blood oxygen saturation, temperature, electrocardiographic signal, atmospheric pressure, number of steps (pedometer) and concentration of compounds present in sweat. As the electrolyte composition in sweat and blood are related, the relevant analysis of sweat can provide a wealth of information on the physical and chemical condition of our body. SportBand is in the form of a belt strap, worn on the chest where the most intensive sweating occurs. During the exercise, SportBand collects sweat and measures the concentration of chlorides, magnesium, potassium, calcium, phosphorus, lactate, urea and glucose. Moreover, it is possible to measure the pH. The tool consists of a single-use polymer insert which can detect, i.e. specific electrolyte concentration. Our biodevice is connected with the smartphone application where it is possible to check all of the collected information along with the description of helpful tips about physical activity and hydration. The presented device is cheap and gives good quality results comparing with the reference methods. It is a promising method for future health monitoring systems not only for the analysis of sweat. The presentation will follow issues like (i) how the concentration of sweat ingredients and other mentioned factors affect our body, (ii) how to measure all of the mentioned factors and (iii) how to use our application.
Nr: 4
Title:

Velo-CT: A Collaborative 3D Racing Game as Incentive for Medical Image Annotation
Authors:

Andreas Maier, Richin Sukesh, Sai Kishore, Srividhya Sathya Narayanan, Vineet Vinay Bhombore, Weilin Fu and Florian Kordon
Abstract: INTRODUCTION. Recent advances in algorithmic processing of medical data are largely based on the use of deep learning algorithms that rely on large amounts of annotated data [1]. To cope with the often complex and time-consuming nature of data acquisition and annotation, data donation and distributed crowd-labeling are an appealing approach [2]. In this paper, we follow this idea and present a single/multi-player mobile game which aids in the annotation process of CT or MR scan data. METHODS. To provide users with an exciting and rewarding annotation experience, our game "Velo-CT" integrates labeling of 3D volume slices into a classical racing game. This is achieved by using data annotation as a key to unlock more advanced racing tracks in the game. To this end, the game has two game modes: 1) Track Racing, and 2) Image Annotation. In “Track Racing”, players compete by steering a vehicle through a racetrack. The main task is to overtake the opponents and try to remain first in position until the end of the race. Checkpoints along the track allow additional guidance if players accidentally leave the racetrack. To provide a more challenging and diverse experience, respawning pick-ups can be found at random locations along the track in the form of floating CT scanners. Such pick-up can either be a speed multiplier (multiplies car speed by 1.2) or a size enhancer (increases the car size by 1.5 to threat opponents and impede their view). The racetracks are modeled based on the segmentation contour of clinical CT data, e.g. showing the respiratory region. For slice segmentation we use a simple K-means classifier. Annotation points earned after beating a track allow the player to unlock new data for the second game mode “Image Annotation”. This is where the medical expertise of the player comes into play. Based on the current progress, the player is presented several medical images which may contain abnormalities. If an abnormality is spotted, the precise location can be marked by a finger tap on the area. The more images a player annotates, the more points can be earned to unlock a new track – and eventually a more intricate anatomy to race on. RESULTS. Our game was prototyped in Unity3D [3]. As a first anatomy, respiratory tracks based on the “Low Dose CT Grand Challenge” data were constructed. First experiences by players of different gaming background were reported for the single player setting. Users report a fast-paced racing experience with a healthy amount of challenge provided by pick-ups and competitive AI opponents. The organ-aligned layouts result in tracks similar in shape, which should be addressed in the future by adding multiple anatomical regions. CONCLUSIONS. Velo-CT provides an intuitive and extendable integration of the often tedious data annotation process for 3D medical data into a rewarding gaming context. The prototype is available for free download and serves as a basis for further extensions to related game modes.
Nr: 5
Title:

ScanRacer: Fast Track Annotation of Segmentation Masks in Volumetric Medical Images
Authors:

Andreas Maier, Jessica Diehm, Robert Hermann, Tobias Pertlwieser, Wei Cheng, Weilin Fu and Florian Kordon
Abstract: Introduction. With the rise of deep learning [1], we see a dramatic need of curated and annotated medical image data. In particular, in volumetric images, such annotations are extremely costly, as each slice has to be outlined individually to generate ground truth for the training of deep learning algorithms. Recently, a new initiative was founded that aims at encouraging patients to donate their medical image data [2]. In particular, this initiative also asks for permission to crowd-source the data annotation. This forms a basis to generate sufficient data for large-scale training of deep learning algorithms in medical image analysis. Methods. In order to encourage users to perform annotations, we explore gamification. In particular, we selected the setup of a racing game to generate an exciting user experience. The main idea is that the user is driving a race car across a volumetric image slice. In order to create tracks automatically, a simple segmentation method is required. For the first experiments, we chose thresholding to separate fore- and background. Then, we select the largest connected component and perform image processing to automatically extract a closed edge contour. As guidance, checkpoints are spread over the preliminary organ outline to guide the player in equidistant steps along the extracted contour. Based on edge detection [3], a score consisting of accumulated edge pixels is determined automatically as feedback for the player. In order to create a more challenging game experience, additional moving obstacles were added to the course. During driving the player creates a closed contour that is then transmitted from the game client to the server’s database. Results. The game was implemented in Unity3D [4] and released as Android APK Installer (https://www.medicaldatadonors.org/index.php/scan-racer/). We chose Google Firebase to store segmentation results on the server (https://firebase.google.com). ScanRacer creates a challenging, yet rewarding experience for the user. Experienced players are able to create segmentation contours close to the correct segmentation outline. Yet detailed annotations still pose a challenge in this setup which will be addressed by the addition of further game elements. In contrast to many other serious games, the fun component of ScanRacer is very high as reported by test players. A gameplay teaser video demonstrates this in detail. (https://www.youtube.com/watch?v=JNmEGLCyf6w)
Nr: 7
Title:

Skeletal Differences among Fed and Starved Fish (Sparus aurata) Determined by Computed Tomography
Authors:

Diana C. Ceballos Francisco, Nuria Garcia Carrillo, Francisco Javier Pardo Fernández, Alberto Cuesta and María Ángeles Esteban Abad
Abstract: Many fish species are exposed to starvation or restricted food intake in particular phases of their life cycles and during this period, they have to direct energy reserves from growth to the support of vital processes, which triggers metabolic changes in many tissues including bones. In this work, we study the possible changes at skeletal level between fed fish and fish subjected to 60 days of starvation through computerized tomography using an Albira SPECT/PET/CT preclinical-scanner. Image analysis and measurements were performed using the Carestream Molecular Imaging Albira CT system in conjunction with Pmod and Amide packages. Boxes ROIs were drawn within the density range previously determined for fish bone (from 200 HU) in Amide software. Results reveal that the mean density value for fish bone was 350 HU for both fed and starved fish; however, the volume occupied by the bone related to the whole body volume in fed fish was higher when compared to starved fish. Related to the densest structure in fish body (otoliths) the volume occupied was similar for both fish groups. These data demonstrate the feasibility of CT to evaluate fish body.
Nr: 18
Title:

HeartPole: A Transparent Task for Reinforcement Learning in Healthcare
Authors:

Vadim Liventsev, Alexandre Simon, Aki Harma and Milan Petkovic
Abstract: Reinforcement learning in Healthcare is an emergent field that has created a demand for patient simulators like GYMIC - a black box neural model trained on MIMIC III dataset that predicts health outcomes of clinical decisions and can be used for training clinical decision-making models. We introduce a patient simulator inspired by CartPole that trades clinical accuracy off for simplicity and transparency, while still being non-trivial to solve.
Nr: 22
Title:

Learning Health System Framework to Continuously Learn from Successes and Failures in Innovation and Accelerate the Use of Health Data to Iteratively Produce New Knowledge to Improve Clinical Care
Authors:

Joanne Enticott, Angela Melder, Alison Johnson, Angela Jones, Tim Shaw, Wendy Keech, Jim Buttery and Helena Teede
Abstract: Background: Information and communication technologies alone do not drive healthcare improvement. System-level approaches are needed as in a Learning Health System (LHS) to operationalise and convert routinely collected health data into useful information to improve decision making and quality care. Here we aim to outline the process and outcomes of developing the Learning Health System framework and outline government funded pilots currently underway in Australia. Methods: This Learning Health System framework development occurred in the context of an Academic Health Science Centre (AHSC) and national alliance, where AHSCs are established as Research Translation Centres in Australia. Following a national priority setting process, we partnered extensively and led the co-design of a Learning Health System framework, applying robust methods across governance, stakeholder engagement, systematic literature review, qualitative research and workshops. Results: The co-design process resulted in bringing together multidisciplinary stakeholders including community, clinicians, academics, administrators and industry; and generated the collective vision of ‘Learning together for better health’. Governance was established and resources acquired through the Australian Government Medical Research Future Fund. The resultant Learning Health System framework aims to integrate stakeholder problems and priorities, research and best practice evidence, data analysis and benchmarking, and implementation and improvement. The system-level approach takes practice to data, integrates best evidence to guide practice, analyses data to generate new knowledge and delivers implementation to take knowledge into practice, embedded within healthcare. This is aligned with a shared vision, robust governance and appropriate infrastructure and multidisciplinary expertise with effective streamlined systems and processes. The framework comprises four key sources of evidence, with each represented as a quadrant of a LHS cycle: • Stakeholder‘s evidence - from end user problems and priorities • Research evidence- from primary research, evidence synthesis and guidelines • Data evidence - from data analysis, including artificial intelligence • Implementation evidence - integrating rigorous implementation research into pragmatic healthcare improvement Each quadrant of evidence is vital to capture, identify and address health service and community priorities and emergent challenges and needs to be integrated to create and operationalise the Learning Health System as an iterative systems level intervention to deliver health impact. Conclusion: The Learning Health System provides a framework for continuous learning from successes and failures in innovation to accelerate the use of health data to iteratively produce new knowledge to improve clinical care and health outcomes. We are now working across partner organisations and with Government and other Centres to pilot and iteratively learn together for better health. Currently there are initiatives using the LHS framework in the four biggest States in Australia. These initiatives utilise big data, networking, graphical interfaces/dashboards, data mining, machine learning, pattern recognition and intelligent decision support systems. Applying the LHS framework to each stage of these innovations will support optimal development, implementation and uptake at pace and scale.
Nr: 29
Title:

Development of a Biomedical Signal-based Driver Sleepiness Detection System: A Supervised Machine Learning Approach
Authors:

Md Mahmudul Hasan, Christopher Watling, Gregoire Larue and Mark King
Abstract: Sleep is a vital part of human life and the deficit if sleep causes impaired cognitive processing in daily activities. Sleep loss is one of the main reasons for drowsiness behind the wheel, which is considered as a challenging issue in Australia and worldwide, resulting in a massive amount of fatalities every year. Though there are several modalities for sleepiness detection, the biomedical signal based measures have been proven to be the most reliable and susceptible to sleepiness. Among all the physiological measures of sleepiness, the Electroencephalography (EEG), Electrooculography (EOG) and Electrocardiography (ECG) signals have been validated as the most reliable measures, which are sensitive to the subjective sleepiness levels. There are several endeavours of using EEG, EOG or ECG signals individually for sleepiness detection. Still, most of the approaches are using neither any feature selection techniques nor the best possible combination of the biosignals. As a result, these individual approaches have several shortfalls, such as, achieving a higher gap between sensitivity and specificity or being not well-generalised; consequently, those approaches do not apply to the real-world driver sleepiness detection. Therefore, this study was performed on 35 participants and investigated the feasibility of using the possible combinations of EEG, EOG and ECG based sleepiness detection using two feature selection techniques, namely ANOVA F Test and Correlation coefficient based feature selection. Using the selected features and considering the subjective sleepiness levels as a ground truth, four traditional supervising machine learning models were developed, namely, K-Nearest Neighbours (KNN), Support Vector Machines (SVM), Random Forest Classifier (RFC) and Artificial Neural Network (ANN). The sensitivity, specificity accuracy and area under the receiver operating characteristic curve were utilised to evaluate the performance of the singular and hybrid signal based sleepiness detection system. While comparing the performance of the individual signal-based approaches with the combined biosignal based hybrid approaches, the hybrid biosignal based approaches showed better performance for all the classifiers. The comparison of the performance of the classifier models indicates that the ANN-based classifier outperforms the other classifiers with an overall accuracy of 83.5% when using the triple combination of EEG, EOG and ECG. The detection system with the combination of the three signals also occupies 85.9% of the area under the receiver operating characteristic (ROC) curve, which shows the best compromise between sensitivity and specificity of the sleepiness detection system.
Nr: 33
Title:

Higher-order Kinematic Analysis of Speech Movement Data
Authors:

Stephan R. Kuberski and Adamantios I. Gafos
Abstract: We present techniques and results of a higher-order kinematic analysis of speech movement data registered by contemporary Electromagnetic Articulography (EMA). In particular, we first demonstrate the applicability of a well-established in the human movement field (but in speech rarely used) spline-smoothing approach and illustrate its superiority over traditional signal representations in EMA research. Second, using an heptic spline-smoothing approach, we reveal a so far unknown set of acceleration-based kinematic relations in data of repetitive speech. This set of empirical relations is finally shown to be the theoretical consequence of what a linguist considers to be the dynamical standard model of speech.
Nr: 33
Title:

Higher-order Kinematic Analysis of Speech Movement Data
Authors:

Stephan R. Kuberski and Adamantios I. Gafos
Abstract: We present techniques and results of a higher-order kinematic analysis of speech movement data registered by contemporary Electromagnetic Articulography (EMA). In particular, we first demonstrate the applicability of a well-established in the human movement field (but in speech rarely used) spline-smoothing approach and illustrate its superiority over traditional signal representations in EMA research. Second, using an heptic spline-smoothing approach, we reveal a so far unknown set of acceleration-based kinematic relations in data of repetitive speech. This set of empirical relations is finally shown to be the theoretical consequence of what a linguist considers to be the dynamical standard model of speech.
Nr: 41
Title:

Alignment of Multimodal Images of Extracellular Vesicles
Authors:

Hannah Janout, Boris Buchroithner, Andreas Haghofer, Stephan Winkler and Jaroslaw Jacak
Abstract: Extracellular vesicles (EV) play an important physiological and pathophysiological role and can be used as natural drug carriers, making them an essential asset in several medical fields. Despite their importance and potential in the medical fields, it is impossible to analyze a single population of EVs with the currently available methods, even if the standardization of quality is essential. At the University of Applied Sciences Upper Austria (FH OO), we are developing a new pipeline for ¨ analysis and quality assessment of EVs through multimodal imaging. EVs are applied to a medium with a visible grid structure and accumulate there. These EVs are modified to carry green fluorescent proteins (GFP), which become visible under a fluorescence microscope (FM). The medium’s EVs are measured once with a FM, measuring the amount of EVs with GFPs, and an atomic-force microscope (AFM), providing the total EV count and mechanical information such as their size and firmness. Additionally, the EVs on FM images are extracted and saved as LOC files through software developed at FH OO, Linz, named ¨ 2CALM (Mayr S., 2020). The recording of images through AFM is more time consuming and complicated than using FM; therefore, only a specific ROI on the medium is recorded with AFM. The mustering of data results in images with different formats, sizes, resolutions, and rotation. Due to the considerable difference in the image’s properties and the partially fuzzy images of the grid structure, standard registration algorithms like SIFT or SURF are not feasible. Thus, a unique processing pipeline and a new way to align FM and AFM is required. The grid structure can be enhanced through preprocessing and used as an anchor point for alignment. The contrast is enhanced in the first step, bringing the grid and EVs more into the foreground. Afterward, a median filter is applied to remove most of the EVs and noise. As the grid structure is more prominent on the AFM image, the FM image requires several more preprocessing steps. On the filtered image, all tubular structures are marked and analyzed through connected component analysis. Thus, only the structures with the largest length and thickness remain, resulting in the grid’s extraction. Afterward, both images are converted to binary, and their grids lines are found with a standard hough line algorithm. As not all lines can be detected perfectly, the grid structure is reconstructed based on the acquired lines. From the complete grid, intersections are calculated. These are used as features to calculate an affine translation, returning the rotation and scaling needed to align the AFM and FM images correctly. Since the LOC data derives from the FM image, these are aligned with the calculated translation. With the correctly aligned images, EVs are extracted from AFM scan data by applying a newly implemented filter, and a comparison can be made. This pipeline has shown excellent results in the analysis of the current data, aligning the images with an inaccuracy of 1-3px. Future steps include improving this pipeline, comparing the quantity of EVs, and analyzing the additional information contained in AFM scan data. Thus, we can analyze the accumulated EVs and determine their quality.
Nr: 42
Title:

Challenges and Benefits of Information Communication Technologies Used by Health Care Aides in Caring for Persons Living with Dementia
Authors:

Antonio M. Cruz, Noelannah Neubauer, Samantha Marshall, Lauren McLennan, Peyman Azad-Khaneghah and Lili Liu
Abstract: Information community technologies (ICTs) have the potential to improve care for people living with dementia and support their care partners. While ICTs such as smartphones are increasingly common for use to by health service providers in their daily workflow, there is little evidence on the benefits of smartphones in supporting dementia in the community. In Alberta, Canada, health care aides, or unlicensed service providers (also called personal support workers or personal care attendants) comprise the second largest workforce, next to nurses, that provide care to older adults. Health care aides provide one-on-one care to our vulnerable older adults in everyday tasks such as self-care, medication management, and social interaction. Our 2012 study showed that health care aides welcomed an opportunity to incorporate mobile technologies to address issues in their workflow, including communication, documentation, scheduling, and safety. The purpose of this scoping review was to examine the range and extent of ICTs use by health care aides to manage and coordinate the care-delivery workflow for clients living with dementia. We followed the PRISMA methodology for systematic literature reviews. Our database sources were Medline EMBASE, CINAHL, and Scopus. Inclusion criteria were studies that: (1) reported the use of website, application, mobile applications or software (2) intended for documentation of care, management of care, organizing care for (3) person with special needs (e.g., person requiring health related personal care, communication, activity tracking such as reminder, follow up), that allowed at least the interaction between users including family or informal (unpaid) caregivers and health care aides, licensed practical nurses, registered nurses, case managers, or other allied health care professionals, (4) were published in English. The search identified 1,323 studies of which 40 were included for analyses. Six different types of ICTs were used by health care aides to manage and coordinate the care of people with dementia., e.g., Electronic health record for home care (22.5%, 9/40), facilitate client assessment, care planning and evaluation (22.5%, 9/40), improve everyday work or patient outcomes (30%, 12/40), communication (12.5%, 5/40), telehealth (7.5%, 3/40). A mixed use of ICT applications was found in two cases (5%, 2/40). In 60% of cases, the health care aides used mobile applications installed in smartphones or tablets to manage the workflow. Overall, 18 barriers were found (lengthy installation requiring trained installer, complex licensing and purchasing processes), 33 challenges (issues or lack of integration with other systems 1(10%), difficulties with scaled or customized implementation, other technical problems). There were 45 benefits (improvement of coordination of care, improved accessibility and communication, prevented and reduced errors, reduced costs and service consumption, reduced repetitive actions, and simplified procedures) in adopting ICTs by health care aides. Health care aides provide one-on-one care to our vulnerable older adults in everyday tasks. It makes a difference if older adults can remain in their homes. Despite challenges, the literature identifies a greater number of benefits to the workflow and quality of care by health care aides for their clients living with dementia.
Nr: 43
Title:

Accounting for Label Uncertainty in Deep Neural Networks for Detection of Breast Tumor in Hyperspectral Images
Authors:

Naomi de Kruif, Lynn-Jade Jong, Veronika Cheplygina, Behdad Dashtbozorg and Theo M. Ruers
Abstract: In 10-30% of breast-conserving surgeries (BSC) some malignant tissue is left behind in patients, which compromises tumor-free survival rate. In these patients, additional treatment such as a re-excision or boost radiotherapy is required to clear residual malignant tissue, which negatively affects the cosmetic outcome and patient’s quality of life. Currently, histopathological analysis is the gold standard for assessing surgical resection margins of breast tissue. However, the examination may take up to one week, thus no feedback can be given to the surgeon during surgery. Hyperspectral imaging (HSI) has shown great potential as an intraoperative margin assessment technique since this novel imaging technique can image the entire resection surface fast. However, one main challenge is to develop an accurate network to classify the tissue type of the resection surface during surgery due to label uncertainty. Herein we developed a double channel convolutional neural network (DC-CNN) that exploits both spectral and spatial features for the tissue type classification of breast tissue slices. The DC-CNN network contains two separate channels of which the 1st channel extracts spectral features and the 2nd channel spatial features. Both feature sets are fed to a fully connected layer for the classification. On top of that, we also propose a method to account for label uncertainty at the tissue transition locations. The hyperspectral data (400-1700 nm) was obtained from the tissue slices after gross sectioning of the resection breast specimens. The histopathology (H&E) results were registered with the HSI images to determine the ground truth labels for each pixel. However, the labels at tissue transition areas are less reliable since these pixels represent volumes that may contain a mixture of different tissue types and these pixels might have an incorrect label due to registration inaccuracies of the H&E and HSI images. To address the uncertainty of labels at transition pixels during training, a new loss function is defined. This loss function reduces the effect of label uncertainty by excluding pixels based on their distance to the transition border. The exclusion distance criterion was defined for each tissue type separately, such that it compensated for the imbalanced labels. For invasive carcinoma (IC), ductal carcinoma in situ (DCIS), connective, and fat pixels with a distance lower or equal to 2, 1, 2, and 6 respectively were chosen as pixels with uncertain labels and were excluded during training. The HSI dataset contains images from 41 patients which were randomly split into the training set and test set. By excluding pixels at the transition border of HSI image, we prepared a CERTAIN test set. The network was twice trained and validated on the training set, with and without accounting for the label uncertainty. After hyperparameter tuning, both networks were evaluated on the CERTAIN test set. The results show that by accounting for pixel uncertainty during training, the performance significantly improved. For the discrimination of tumor (i.e. IC & DCIS) from healthy tissue (i.e. fat & connective), the Matthews correlation coefficient and accuracy increased from 0.70 to 0.91 and 0.90 to 0.96, respectively. The proposed method reduces the effects of label uncertainty and significantly improves classifier performance.
Nr: 44
Title:

Using Deep Learning to Predict Chromosomal Abnormalities from Embryo Microscope Images in Preimplantation Genetic Screening
Authors:

Liou Kai Jhong, Cheng-Wei Wang and Emily C. Su
Abstract: Objective: Single embryo transfer (SET) is the current trend in artificial reproduction technology. SET avoids the higher risk of disease caused by multiple pregnancies to the mother and fetus, and the resulting public health issue. However, under the SET model, how to ensure the quality of embryos to improve the live birth rate and reduce the psychological pressure caused by failure to the mother is very important. Commonly used embryo quality assessment methods include morphological assessment and preimplantation genetic screening (PGS). PGS is the most reliable method for assessing embryo chromosomal characteristics, but it still uses cells through biopsy. There is controversy about whether it will harm the embryo. This study tries uses embryo images to predict the chromosome detection results of PGS. The predicted results assist doctors in judging the quality of embryos and reduce the chance of damage to the embryo and the high cost caused by PGS. Method: This study uses the Day 5 images of 1220 embryo(610 Aneuploidy and 610 euploidy) with PGS results accumulated by the Department of Obstetrics and Gynecology of Taipei Medical University Hospital.The model was trained using embryo image by different convolutional neural networks(CNN) architecture(Inception, InceptionResnet, ResNet152V2, etc.) and parameters.The performance of predictive was determined by evaluating accuracy, sensitivity, specificity and area under a curve(AUC) with 5-fold validation. Results: The final model of this study is trained by ResNet-152V2 model, optimizer Adam, learning rate 0.0001, Dropout rate 0.4, batch size of 64,input shap 300 x 300, random rotation and random horizontal flip.Accuracy of this AI model was 77%, sensitivity 77%, specificity 78% and AUC 0.85. This model uses 278 untrained Aneuploidy images as test dataset, the accuracy of test dataset is 77%. Conclusions: At the beginning of this experiment, only 800 image were used, and the results obtained were accuracy 69%, AUC 0.76, sensitivity 71% and specificity 67%. With the collection of data up to 1220 pictures and parameter adjustments, the model’s performance has been significantly improved, so it is believed that as the amount of data accumulates and technology advances, the performance of the forecast will become better.
Nr: 45
Title:

Postsurgical I-123 Thyroid SPECT/CT Imaging using a Custom-made Neck-thyroid Phantom with Small Sizes of Thyroid Remnants
Authors:

Anastasia C. Hadjiconstanti, Konstantinos Michael, Demetris Kaolis, Theodoros Leontiou, Antonios Lontos, George Demosthenous, Savvas Frangos, Diogenis Kyprianou and Yiannis Parpottas
Abstract: Background: The treatment of differentiated thyroid cancer typically involves surgical removal of the whole or the largest part of the thyroid gland, and subsequent radioiodine therapy. Postsurgical thyroid imaging can provide further information on the presence of thyroid remnants and/or metastasis. Phantoms can be used in postsurgical SPECT/CT thyroid imaging to provide information about the presence and sizes of remnants and then to accurately determine the therapeutic doses for ablation. The objective of this study was to validate a custom-made neck-thyroid phantom with small sizes of thyroid remnants in postsurgical I-123 thyroid SPECT/CT imaging. Materials and Methods: The custom-made neck phantom encloses human-sized trachea, oesophagus and cervical spine. At the correct anatomical position, a custom-made removable thyroid-remnant section is also attached at the neck phantom. The hollow cavity of the neck phantom can be filled with water to simulate the soft tissue. Diluted radiopharmaceutical can be injected within this cavity to simulate various thyroid remnant-to-background activity ratios. For this study, the removable thyroid-remnant section enclosed hollow cavities of 1.5 and 3 mL to simulate thyroid remnants after thyroidectomy. However, different removable thyroid-remnant sections with various sizes of thyroid remnants, at any clinically relevance areas, can be attached at the neck phantom. Diluted radiopharmaceuticals can be injected within the thyroid remnants. I-123 SPECT acquisitions were performed with Low-Energy General Purpose (HEGP) collimators in 180° (H-mode) orientation. Data were acquired in 60 projections over 180° of rotation, covering an angular range of 360°. The acquisition time per projection was set to 35 sec. An ±10% energy window was set around the 159 keV I-123 photopeak. Acquisitions were performed for different I-123 administered activities within the thyroid remnants. This range of activities can be administered for diagnostic thyroid SPECT imaging, following an almost immediate acquisition. Images were scatter corrected before reconstruction. SPECT data were reconstructed using the OSEM algorithm. The image matrix size was 128x128 with a pixel size of 4.42 mm. A Butterworth filter was applied to the reconstructed images. CT scans were also acquired to correct the SPECT images for attenuation. First, the response of the SPECT modality was investigated from the total measured counts in each remnant with respect to the administered activity. Second, the sensitivity (counts/(sec×activity)) of each remnant from the custom-made phantom was compared with the corresponding sensitivity of equal sizes of remnants from a modified RS-542 commercial neck-thyroid phantom, in which radiopharmaceutical can be injected only within the remnants. Results: A linear response was concluded for the SPECT modality when investigating the total measured counts in each remnant with respect to the administered activity. Also, the measured sensitivity was comparable among the custom-made and the commercial phantoms in the whole range of activities. Conclusions: This custom-made phantom with small sizes of thyroid remnants, that can also simulate background activities, can be used to evaluate postsurgical thyroid SPECT/CT imaging. Acknowledgments: This study was co-funded by the European Regional Development Fund and the Republic of Cyprus through the Research and Innovation Foundation (Project: EXCELLENCE/1216/0088).
Nr: 46
Title:

GAP: A Genome Annotation Pipeline for Building Exome Annotation
Authors:

Ruoxian Huang
Abstract: Next-Generation Sequencing (NGS) technology (e.g. exome sequencing or whole genome sequencing) is becoming increasingly popular and powerful in identifying genetic variants or mutations, including ones associated with rare diseases and disorders. However, annotating each identified variant requires the right format of input data. It is imperative to design a generalized bioinformatic method, which can take input gene lists and a whole-genome annotation table to dissect necessary genomic contents associated with gene/variant annotation for facilitating further genomic variant characterization. In this study, we have developed a pipeline that can effectively take a list of genes, by processing the University of California Santa Cruz (UCSC) genome annotation table and report the corresponding exon indices with coding information marked for each gene. The generated exon annotation file can facilitate variant function classification. Specifically, we downloaded the Known Gene annotation table from the UCSC Genome Browser and then created a customized Python script to take the input data table and generate the output file in the context of characterizing exons. Our python script annotates the start and end features (coding information) for all exons of each gene including intron, coding region (CDS), upstream, downstream region, and untranslated regions (UTRs). The generated output file for exon annotation from our pipeline is populated with newly added columns that contain the detailed annotation for each gene structure at the exon level, including Coding Start Information (CSI), Coding End information (CEI), Exon Start (ES), Exon End (EE), CDS Start (CS), and CDS End (CE). We have examined our program with DDX3X, a multi-isoform gene associated with intellectual disability (ID) diseases. Our pipeline confirmed nine isoforms for this gene and reported the exon annotation information at the isoform level. The output data generated by our pipeline can be used as a reference file for mapping identified variants to the genomic location and classifies variants by regions.
Nr: 47
Title:

Image Quality from a Quantitative Analysis of I-131 and I-123 Scatter Corrected SPECT/CT Images using a Phantom with Thyroid Remnants
Authors:

Konstantinos Michael, Anastasia Hadjiconstanti, Elena Ttofi, Antonios Lontos, Yiannis Roussakis, Demetris Kaolis, Savvas Frangos and Yiannis Parpottas
Abstract: Background: Postsurgical SPECT/CT thyroid imaging with I-131 or I-123 is performed to evaluate the disease, localize and estimate sizes of remnants, before radioiodine therapy. This study assessed the image quality of I-131 and I-123 SPECT/CT images using a phantom with thyroid remnants. Materials and Methods: Measurements were performed using a custom-made anthropomorphic neck-phantom with a removable section of 1.5- and 3 mL thyroid remnants. Diluted radiopharmaceutical can be injected within the remnants and also within the neck cavity to simulate various thyroid remnant-to-background activity ratios. The SPECT/CT images were acquired in H-mode. The High-Energy General-Purpose collimators were used for the I-131 and the Low-Energy High Resolution for the I-123 acquisitions, respectively. Data were acquired in 60 projections over 180° of rotation (covering an angular range of 360°), 35 sec/projection. A ± 10% energy window was set around the 364 keV I-131 and the 159 keV I-123 photopeaks, respectively. SPECT data were reconstructed using the OSEM algorithm. The image matrix size was 128x128 with a pixel size of 4.42 mm. A Butterworth filter was applied to the reconstructed images. CT scans were also acquired to correct the images for attenuation. The dual-energy window (DEW) method of Xeleris workstation was also applied on images for scatter correction. The first set of acquisitions was performed without background activity. The administered activities within the remnants were from 2.5 to 40 MBq for the I-131 measurements and from 0.5 to 6 MBq for the I-123 measurements, respectively. Note that, the administered activities for the I-123 acquisitions were lower than the I-131 ones due to the high counting rate of I-123. The first part of acquisitions was used to study the modality response and the counting rate. The second set of acquisitions were performed with I-123 and I-131 for different background activities within the neck cavity to achieve different remnant-to-background activity ratios (5% and10%). In these acquisitions, the administered activity within the remnants was always 0.37 MBq/mL. These acquisitions were used to calculate the contrast-to-noise (CNR) and signal-to-noise (SNR), and consequently to decide the image quality. Results: The I-131 and I-123 counts with respect to the administered activity showed a linear response of the SPECT modality. The rate for the I-123 acquisitions was much higher than I-131. The average sensitivity (counts per second per activity) for I-123 was 3.5 times higher than the corresponding one for I-131. Τhe CNR and SNR values are significantly higher for the I-123 than the I-131 images. Both CNR and SNR values are significantly better for the 3 mL than the 1.5 mL remnants, especially for the 10% background activity. When DEW scatter correction was applied, the CNR and SNR values were increased for both I-123 and I-131. This increase was more profound in the case of I-131. Conclusions: I-123 provides higher counting rate than I-131, even in lower administered activities. The CNR and SNR values indicate that the quality of I-123 images is higher than the I-131 images. DEW scatter correction also improved the quality of both I-123 and I-131 images, especially the I-131 images. Acknowledgments: This study was co-funded by the European Regional Development Fund and the Republic of Cyprus through the Research and Innovation Foundation (Project: EXCELLENCE/1216/0088).
Nr: 48
Title:

Two Scatter Correction Methods on I-131 and I-123 SPECT/CT Images using a Custom-made Phantom with Thyroid Remnants: Image Evaluation from Physicians
Authors:

Ioannis Petrou, Anastasia C. Hadjiconstanti, Christoforos Panagidis, Savvas Frangos, Konstantinos Michael and Yiannis Parpottas
Abstract: Background: Differentiated thyroid cancer treatment typically involves surgical removal of the whole or the largest part of the thyroid gland and a subsequent radioiodine therapy. Postsurgical nuclear imaging with I-131 or I-123 can provide further information on the presence of thyroid remnants and/or metastasis. The diagnostic accuracy of the thyroid remnants is important for the implementation of individualized treatment for remnant ablation. In this study, the dual energy window (DEW) and the triple energy window (TEW) scatter correction methods on I-131 and I-123 SPECT/CT images of a custom-made phantom with thyroid remnants were visually evaluated by three experienced nuclear medicine physicians. Materials and Methods: SPECT/CT acquisitions were performed using a custom-made phantom, which encloses trachea, oesophagus, cervical spine and a removable section with thyroid remnants. For this study, remnants of 1.5- and 3 mL were inserted, at clinically relevance areas, to simulate thyroid remnants after thyroidectomy. Radiopharmaceuticals can be injected within the hollow cavities of the remnants (target) and the neck (background). The clinical protocols were followed for the acquisitions. The High-Energy General-Purpose collimators were used for the I-131 acquisitions and the Low-Energy High-Resolution (LEHR) collimators for the I-123 acquisitions, in 180° (H-mode) orientation. Data were acquired in 60 projections over 180° of rotation, covering an angular range of 360°, with 35 sec/projection. A ± 10% energy window was centered over the 364 keV photopeak of I-131 and the 159 keV photopeak of I-123, respectively. For the DEW and TEW scatter correction of I-131 and I-123 acquisitions, the width of all scatter energy window was set to 5 keV. SPECT data were reconstructed using the ordered-subset expectation-maximization (OSEM) algorithm. The image matrix size was 128x128 with a pixel size of 4.42 mm. A Butterworth filter was applied to the reconstructed images. CT scans were acquired to correct for attenuation the non-scattered corrected (NSC), DEW and TEW scatter corrected SPECT data. The first part of acquisitions was performed for different administered activities of I-131 and I-123 within the 1.5- and 3 mL remnants. The second part of acquisitions was performed for different remnant-to-background activity ratios (5%, 10%) of I-131 and I-123. Results: The physicians were more confidence to evaluate the volume of each remnant from the TEW scatter corrected SPECT/CT images than the DEW ones, especially when reading the images with background activity. This is due to the higher quality of TEW images. However, both scatter correction methods improved the quality of I-131 and I-123 images. Comparing the I-123 and I-131 images, the physicians considered that the quality of the I-123 images is higher than the I-131 ones, even with and without background activity. Conclusions: The diagnostic postsurgical thyroid SPECT/CT images of a custom-made phantom were evaluated by three experienced nuclear medicine physicians. The quality of the TEW images is higher than the DEW ones for both I-131 and I-123. In all cases, the quality of I-123 images is superior than the I-131 ones. Acknowledgments: This study was co-funded by the European Regional Development Fund and the Republic of Cyprus through the Research and Innovation Foundation (Project: EXCELLENCE /1216/0088).
Nr: 49
Title:

Inferior Artifact on Myocardial Perfusion SPECT Imaging Due to Liver
Authors:

Sotiris Panagi, Demetris Kaolis, Diogenis Kyprianou, Costas Kyriacou, Ioannis Petrou and Yiannis Parpottas
Abstract: Background: Coronary artery disease (CAD) is the most common form of heart disease worldwide. Myocardial perfusion imaging (MPI) has been proven to be able to assess the functional significance of a coronary artery stenosis. Whereas MPI is a valuable diagnostic imaging process, it is subjected to a variety of artifacts that can limit the performance of the study. Patient-related artifacts in MPI, not originated from the heart, are mainly due to attenuation, thoracic motions, and sub-diaphragmatic activity. Sub-diaphragmatic organs, mainly the liver, present prominent radioactivity which mainly interferes with the adjacent inferior wall of the left ventricle. In this study, we investigated the influence of liver activity on MP during cranio-caudal respiratory motion. Material & Methods: SPECT acquisitions were performed using a phantom assembly of an anthropomorphic thorax phantom with thoracic moving phantoms. The thorax encloses, in the proper anatomical positions, human-sized and shaped thoracic phantoms of: (a) an ECG beating cardiac, (b) inflatable lungs, and (c) a liver. These thoracic phantoms can also oscillate in the cranio-caudal direction for different respiratory amplitudes. Diluted radiopharmaceutical can be injected within the myocardial wall of the left ventricle (LV) of the cardiac phantom and within the liver phantom. Defects of different extent and thickness can be positioned in any LV segments. Acquisitions were performed with a Philips Forte SPECT utilizing two Low-Energy High-Resolution collimators in 90° (L-mode) orientation. Data were acquired in 64 projections over 180° of rotation (covering a range of 360°) on a 64 x 64 matrix for 25 seconds per projection. A 20% energy window was centered over the 140 keV photopeak of Tc99m. Acquisitions were performed for different cases of phantom assembly: (a) two cardiac-liver activity ratios (CLA = 1:0 and 1:0.5), (b) without and with cardiac inferior defect (2x2 cm in extent and 1 cm in thickness), and (c) two cardiac-liver proximities (CLP = 0.5- and 1.5 cm at diastole). The abovementioned acquisitions were performed for a non-moving and a moving, at deep respiration of 2.7 cm oscillatory amplitude, phantom assembly. From the acquired polar maps, the regional uniformity in the inferior area (RU) was calculated. For the abovementioned cases, we investigated the change of regional uniformity, in the inferior area of LV, between the moving and non-moving phantom assembly. Results: In deep respiration, RU was decreased, as expected. This decrease of RU was higher when acquisitions were performed for CLA of 1:0.5 than 1:0. Also, this decrease of RU was higher when acquisitions were performed with the abovementioned inferior defect. In addition to this, in acquisitions where a cardiac defect and liver activity were presented, this decrease of RU was higher for CLP of 1.5 cm than 0.5 cm. Conclusions: Deep respiration produces an inferior artifact due to motion. This artifact can be more profound due to liver activity and, in particular, when an inferior defect is presented. Moreover, this artifact is less intense when the liver is closer to the heart since the liver activity contributes to the inferior LV wall. Acknowledgments: This study was co-funded by the European Regional Development Fund and the Republic of Cyprus through the Research and Innovation Foundation (Project: EXCELLENCE/1216/0085).
Nr: 50
Title:

Motion Correction on Myocardial Perfusion SPECT/CT Due to Liver Activity using a Phantom Assembly with Dynamic Respiratory Phases
Authors:

Irene Polycarpou, George Charitou, Sotiris Panagi, Demetris Kaolis and Yiannis Parpottas
Abstract: In myocardial perfusion SPECT imaging, patient motion may lead to image blurring primarily in the cranio-caudal direction. This can result in misleading displacement of uptake affecting the perceived regional localization in the myocardial walls. Motion induces artifacts, mainly in the anterior and inferior walls, that can be a source of false-positive findings and may affect the ability to detect defects. Furthermore, the activity of the moving heart and liver in the cranio-caudal direction during respiration overlap, increasing the uptake in the adjacent heart inferior wall. Different amplitudes of these motions may result in different spill-over of the liver activity into the adjacent heart walls. This varying liver interaction cause a complex and clinically unpredictable variation in MPI. Correction of this cardiac motion is desirable to reduce these artifacts, and consequently to improve diagnosis. The aim of this study was to correct the cardiac respiratory motion under the influence of the liver activity. Acquisitions were performed using the SPECT/16-slice-CT. For this purpose, a phantom assembly of an anthropomorphic thorax phantom which enclosed thoracic moving phantoms was utilized. The thoracic phantoms are an ECG beating cardiac, inflatable lungs, and a liver. They can oscillate in the cranio-caudal direction for different respiratory amplitudes. The phantom assembly is able to perform different dynamic respiratory phases. Respiration was separated into isochronous static respiratory phases from end-expiration to end-inspiration. A dynamic respiratory phase oscillates between nearby static respiratory phases of the thoracic phantoms. Tc99m SPECT/CT data were acquired with the phantom assembly being at end-expiration (reference) and with the phantom performing each of the three respiratory phases for a deep respiration of a max amplitude of 2.7 cm. These acquisitions were performed with and without liver activity (heart-to-liver activity ratio of 1:1). Scatter and attenuation correction were applied and data were reconstructed with and without motion correction. A non-linear registration algorithm was applied to register each of the reconstructed SPECT dynamic respiratory phases to the reference one and obtain displacement vectors between them. Each dynamic phase was transformed to the reference one and were averaged to create the motion corrected image. Images with and without motion correction were quantitatively compared. The cardiac respiratory motion decreased the uptake in the anterior and inferior walls of the left ventricle. Liver activity interferences with the inferior LV wall increasing this motion-related reduced uptake. Motion corrected images, acquired with and without activity within the liver phantom, have an impact on these artifacts. MPI motion-related artifacts can be corrected with this motion correction algorithm. Interpretation of the motion correction on SPECT MPI may be affected by the inferior wall artifacts introduced from liver activity. Care should be taken when evaluating SPECT MPI studies with breathing abnormalities and increased liver uptake. Acknowledgments: This study was co-funded by the European Regional Development Fund and the Republic of Cyprus through the Research and Innovation Foundation (Project: EXCELLENCE/1216/0085).
Nr: 51
Title:

Development of a Wearable Device for Simultaneous Monitoring of ECG and Thoracic Impedance, Including Respiratory Information
Authors:

Daisuke Goto, Takuya Toyoshi, Masanobu Manno, Shima Okada and Naruhiro Shiozawa
Abstract: Heart rate (HR) and R-R interval (RRI) are used widely not only in sports but also in the medical and healthcare fields for predicting heart failure (HF) and monitoring health conditions. Impedance is obtained by applying micro current to the human body. Especially thoracic impedance includes important cardiorespiratory parameters, such as cardiac output, respiration rate, and amount of ventilation. In the conventional method, a lot of information cannot be obtained without using many sensors or electrodes. However, it is desirable to measure many bio-signals with the minimum number of sensors for the user’s comfort and to reduce running costs.. The objective of this research is to develop a wearable device for simultaneous monitoring ECG and impedance including respiratory information using two dry electrodes. We designed a circuit for simultaneous monitoring ECG and impedance. The dry electrodes were pasted on the garment and electrode positions were the CC5-lead. A stretchable conductive sheet (K3B80S, TOYOBO, Osaka, Japan) was used as the electrode. The subjects were ten healthy males (Age: 21.7 ± 1.3 yrs, Height: 169.3 ± 5.0 cm, Weight: 60 ± 6.6 kg). Each subject wore the developed wearable device with two dry electrodes. For comparing the developed device's measurement accuracy, two Ag/AgCl electrodes (Blue Sensor, SP-50-01) were pasted on the 2-lead induction position, the golden standard measurement method. Sebum was removed using an alcohol wipe before pasted the Ag/AgCl electrodes. To measure respiration, a thermistor (R-101A, NIHONKOHDEN, Tokyo, Japan) was pasted on the nose and mouth of each subject. To compare the biopotential by a conventional method, PolymateV (Miyuki Giken, Tokyo, Japan) was used. The electrocardiographic waveform and the thoracic impedance were measured during the rest sitting position. To confirm the respiration rate, the rate was set to 15 and 30 breaths per minute. The sampling frequency of all devices was recorded at 1 kHz. Measuring the impedance between the dry electrodes of the developed device, an LCR meter (IM3590, Hioki, Nagano, Japan) was used. The electrocardiographic waveform and impedance data from the developed and the reference device were converted into a digital signal by an AD conversion device (PowerLab 16/30, ADInstruments, Sydney, Australia) with 16-bit resolution, and recorded at PC. The average R-R interval (AVNN), the average impedance, and the average respiratory interval were calculated to compare the accuracy of the developed device and the reference device. The AVNN of both devices was coincided (R2 > 0.99) in the ECG measurement. The difference of the average impedance was 5.3 ± 3.8 Ω between both devices and the mean of absolute respiratory interval errors between both measurements were 0.01 seconds in impedance measurement of two tests. The accuracy of the impedance value from the developed device was approximately the same as the reference device. The result of this research shows that the developed device can measure heart rate and thoracic impedance including respiratory rate simultaneously.