Biomedical Engineering

Permanent URI for this collection

http://etd.aau.edu.et/handle/123456789/50

Browse

Now showing 1 - 20 of 60

Investigating and Automating Ethiopia's Radiation-Generating Medical Device Inspection System
(Addis Ababa University, 2022-04) Alemu Nigusse; Masreshaw Demelash (PhD)
Ionizing radiation is energy in the form of wave or particles that have enough energy to remove electrons from the atoms. Since this radiation imposes a deterministic and stochastic effects on human being, the source needs strict regulation to minimize the risks and hazards as low as rea-sonably achievable. In the healthcare sector, the major sources of ionizing radiation are Radiation Generating Devices (RGDs) such as radiotherapy machine, CT-Scan machine, and X-ray machine. Ethiopian Radiation Protection Authority (ERPA) is established to regulate such sources during import, transport, use, and export. However, previous studies revealed the exposure of workers around this practice is above the exposure limit of International Basic Safety Standard. The aim of this thesis is therefore to investigate the gaps around regulation of RGDs and showcase one way to improve the regulatory system. For the investigation, an institutional based cross-sec-tional study was carried out in nine hospitals and one authority (ERPA) from October in 2016 to February in 2017 in Addis Ababa. A total of 102 professionals were purposively selected as re-spondents based on the inclusive criteria, out of which 72 were selected from nine hospitals and the rest 30 were selected from ERPA. The data were collected through structured questionnaires, focused group discussion, document analysis and practical observation. The result indicates the overall mean of respondents who reported that there were regulatory in-spection was 73.8% in ERPA and 65.1% in hospitals; checked out each safety control parameters was 89.8% in ERPA and 80% in hospitals; and checked out each performance indicator parameters was 81.7% in ERPA as well as 66.9% in hospitals. Furthermore, more than 90% (27) of partici-pants, for each item, in ERPA believed that there is no enough professional man power in ERPA; and 90% (27) of respondents disagreed that the expected inspection frequency is specified for each devices; and the overall mean of about 64.9% of participants in ERPA disagreed on each staff management requirements. Moreover, more than half (50.5%) of respondents in hospitals disagreed and strongly disagreed on availability and appliance of each inspection procedure. The finding of this thesis revealed different gaps during regulatory inspection such as missing of RGDs, bypassing safety control and performance indicator parameters; and lack of material and human resource in the staff of ERPA. HTML5, MySQL, PHP, XAMPP, and CSS3 soft- wares were employed to develop a Web-based scheduling and registration system. It is believed that the developed system can be a showcase solution for the problems related to accessing and retrieving, registering & scheduling, deleting & adding, editing, storing, and updating any necessary infor-mation of RGDs that the regulatory body needs to do its activities properly.
Automatic Image Analysis for Bone Age Determination
(Addis Ababa University, 2022-12) Andualem Wube; Dawit Assefa (PhD); Asfaw Atinafu (PhD)
Bone age evaluation is commonly performed through radiological assessment of the skeletal development of the left hand, and then is compared with the chronological age. However, despite the time taking process, such image based automated processing remains incredibly challenging to implement. An accurate method is important to study the development of different wrist and pelvis structures which can predict age effectively. The proposed age determination scheme is composed of two major steps a watershed based segmentation scheme used to generate segmented images which are used as inputs to a deep learning algorithm that is able to determine the age of a given subject. The deep learning scheme utilized the InceptionV3 architecture for its implementation. The model was trained, validated, and tested on radiological hand images acquired from the RSNA database. The model resulted in a mean square error of 4.4 months when compared against the available ground truth information. Overall, the results showed that the proposed bone age determination scheme comes with great promises.
Glaucoma Detection Using Macular OCT Images Based on Deep Convolutional Neural Networks
(Addis Ababa University, 2023-03) Hana Mekonen; Dawit Assefa (PhD); Tesfaye Tadesse (PhD) Clinical Advisor
Glaucoma is a major public health problem as it is the second leading cause of blindness after cataract. Since vision loss due to glaucoma can't be recovered, an early, reliable diagnosis is desirable. Although complete eye examination is recommended for assessment of both structural and functional states of the disease, glaucomatous structural changes precede functional changes. For instance, many studies reported that 25-30% of ganglion cell loss precedes the manifestation of visual field defect and loss of retinal nerve fiber layers (RNFL) occurs approximately six years before any detectable visual field defect. Therefore, the early diagnosis of glaucoma relies on the detection of these structural changes. Recently, classifying glaucomatous images taken from different modalities based on Deep Learning (DL) is increasingly being studied. Most of the researchers, however, relied on images generated from a fundus camera and others on OCT scans taken from the optic nerve head (ONH). Various others relied on specific information derived from the OCT machine itself including thickness and deviation maps of macular and ONH scans, and en-face images. However, the glaucomatous eye can be more effectively detected by analyzing the degeneration of the ganglion cell complex (GCCs) by using original OCT complete scans as input. The current thesis study used deep segmentation models to extract the GCC region which is composed of the retinal nerve fiber layer and ganglion cells with the inner plexiform layer. The study also used Convolutional Neural Network (CNN) based classifiers for detecting glaucomatous pathologies by paying attention to the GCC region of the macula Spectral Domain Optical Coherence Tomograpgy (SD-OCT) scans. The data set utilized for training and validation of the models composed of 1,262 locally acquired macula SD-OCT B-scans (431 non-glaucomatous and 830 glaucomatous) from four different regions of the macula: superior outside, inferior outside, inferior inside and central macula regions. Transfer learning was employed for segmentation as well as classifying the dataset. Deep segmentation models, SegNet, PSPNet, and RAG−Netv2, were employed for segmentation and CNN models namely VGG16, VGG19, and ResNet50 were used for classification purpose. SegNet showed the best performance for retinal layer segmentation with 97.89% accuracy, 87% recall, 87.5% f1-score, 88% precision, 89% mean dice coefficient, and 81% mean_IOU. In terms of classification of glaucomatous and normal images, the best accuracy of 94.3% was obtained using VGG16 computed on the superior outside macula region, with 93.3% precision, 91.7% recall, 91.8% f1-score and 91.7% AUC. The study has demonstrated that using GCC aware deep learning model based on macula B-scans show great promises in accurate screening of glaucoma and suggested that incorporating DL technology into macula SD-OCT for glaucoma assessment has the potential to fill some gaps in current practices and clinical workflow.
Magnetic Resonance Images Based Cervical Cancer Classification Using Convolutional Neural Network
(Addis Ababa University, 2022-06) Kidist Kebede; Dawit Assefa (PhD)
Cervical cancer is one type of cancer which affects the cells of the lower parts of the uterus that connects to the vagina called cervix. According to statistics from the ICO/IARC HPV information center, the yearly estimate of cervical cancer in 2020 was 604, 127 new cases and 341, 831 deaths globally, and the second most common in developing countries. It is also reported that more than 85% of cervical cancer patients are living in low resource setting countries being a major cause of morbidity and mortality. To reduce the death rate, effective screening, diagnosis, staging and treatment should be available. The current thesis study aims to bring automatic staging scheme to be used as a decision support system for cervical cancer treatment and prognosis. Pelvic Magnetic Resonance Images were acquired from St. Paul Hospital Medical Millennium College and Tikur Anbessa Specialized Hospital. Basic image pre-processing was employed on the acquired input images. Then two dimensional convolutional neural network was utilized as an integrated feature extraction and classification scheme. The effect of Network layer variations and important network hyper parameters, including learning rate, number of filters, kernel size and epoch was investigated. The performance of the proposed algorithm in binary (two class) and multi (three and five class) classification were tested and resulted in best classification accuracy of 85%, 68.8% and 56.9% respectively. CNN performance was also compared against two other machine learning approaches, namely Support Vector Machine and K Nearest Neighborhood, where both employed region descriptors as well as Gray level co-occurrence matrix during feature extraction. Results showed that the proposed Neural Network based classification scheme outperforms the two machine learning approaches showing its great promises to assist physicians as a decision support system.
Characterization of Antibubbles Response under Sonication In Terms of Size Distribution and Destruction Threshold for Use in Drug Delivery
(Addis Ababa University, 2023-06) Mahider Yifru; Dawit Assefa (PhD); Michiel Postema (PhD) Co-Advisor
Antibubbles are gas bubbles containing liquid droplets. Due to their acoustic property, they are preferable for different medical applications such as drug delivery and ultrasonic harmonic imaging. In drug delivery, they are used as drug carrier for effective treatment of diseases such as cancer. In order to ensure safe and targeted drug delivery, the behaviors of the drug carriers (antibubbles) need to be studied. The aim of the current research was to study response of antibubbles under sonication and determine destruction threshold which can lead to fragmentation under ultrasound pulse having safe mechanical index (MI) with reduced undesired mechanical and thermal bio-effects. Gray-scale video frames of antibubbles were used for this research, which were generated through laboratory experiment. Matlab and ImageJ platforms were used for effective detection of antibubbles and determination of their respective size before, during and after sonication together with destruction threshold. The proposed algorithm involves extraction of texture features form the video frames based on the local gray level co-occurrence matrix (GLCM) and a watershed scheme for segmentation of the antibubbles. Once antibubbles are detected, their size distribution under ultrasound pulse with varying acoustic pressures was analyzed. The investigation carried out under sonication of low pressure ultrasound waves (0.2 MPa and 0.4 MPa) showed that the antibubbles oscillate repeatedly with series of stable contraction and expansion but will not undergo fragmentation. This property of antibubbles interaction with low pressure ultrasound wave makes them very useful and preferable to be used as a drug carrier in a way of manipulating the antibubble movement through the blood vessel to the target region. In case of antibubble sonication under higher acoustic pressure wave (0.6 MPa and 1 MPa), their oscillations become more asymmetric which finally lead to fragmentation. The MI of 0.6 MPa and 1Mpa ultrasound pulses are 0.6 and 1 respectively, where significant risk of cavitation is considered for MI > 0.7. The destruction threshold analysis showed that antibubbles under acoustic pressure of 0.6 MPa experience fragmentation when their size during maximum expansion is approximately twice that of their initial size intermes of area, i.e. Amax ≈ 2.1 Ai. In the case of 1 MPa acoustic pressure Amax ≈ 2.8 Ai, showing that ultrasound pulse with acoustic pressure of 0.6 MPa is considered safe compared to 1 MPa for use in drug delivery with the capacity to induce fragmentation and effective targeted drug release.
Artificial Intelligence-Based Clinical Assistance Tool for COVID-19 Cases in Ethiopia
(Addis Ababa University, 2022-12) Mastewal Mathiwos; Dawit Assefa (PhD); Robel Kebede (PhD) Co-Advisor
The COVID-19 crisis has had a devastating impact in terms of loss of human life and economic disruption. According to a world health organization report on March 2022, more than 6.1 million people have died worldwide as a result of this pandemic. Early and precise recognition of COVID-19 is key for treating patients and slowing the spread of the pandemic. Several artificial intelligence (AI) based solutions have been developed to facilitate the application of chest X-ray (CXR) imaging and anthropomorphic data for use as a COVID-19 screening tool in resource-limited settings. The current study aims to develop a COVID-19 diagnosis scheme based on a deep learning (DL) approach applied on X-ray image samples collected locally in Ethiopia as well as a severity prediction tool based on a machine learning (ML) approach applied on anthropomorphic data collected from the same patients. The study data for the DL approach consisted of 746 CXR labeled images collected from St. Peter’s Specialized Hospital (SPSH) and Millennium COVID-19 care center (MCCC) in Ethiopia. The samples were composed of images acquired from patients treated for COVID-19 and normal controls. The DL model involved data preprocessing steps applied on the raw CXR images and was designed based on transfer learning approaches. The diagnostic prediction ability of the approach was measured in terms of probability score value and occlusion sensitivity map. The study data for the ML model development consisted of 308 anthropomorphic data (including demographic, comorbidity, and COVID-19 symptoms) collected from the MCCC to check the severity level of the COVID-19 cases and classify them into one of the three classes: moderate, sever or critical. A 5-fold cross-validation approach was used to train two popular ML approaches namely Support Vector Machine (SVM) and K-nearest Neighborhood (KNN) models. The DL method using ResNet-50 architecture achieved best classification performance with a validation accuracy of 94.20 % in accurately classifying COVID-19 and normal cases. The SVM model achieved a better prediction ability than K-NN with an overall accuracy of 89.9 % in predicting the severity status of the COVID-19 cases.
Genetic Algorithm Based Optimized Radiotherapy Patient Scheduling
(Addis Ababa University, 2022-02) Metadel Hailu; Dawit Assefa (PhD)
Radiotherapy is the major means to treat cancer patients. Radiotherapy comprises two phases: pretreatment and treatment on radiation machines. This thesis work focuses on the treatment phase. Treatment consists of multiple, almost daily irradiation appointments, followed by optional imaging and control assignments. The scheduling of radiotherapy appointments is a complex problem due to various medical and scheduling constraints, such as patient category, machine availability, waiting time targets and also due to the size of the problem (i.e., number of machines, facilities and patients). The objective of this thesis is to minimize waiting time and maximize device utilization in) patient’s appointment scheduling. Thus, this thesis presents an optimization algorithm for scheduling of radiotherapy treatments for categorized cancer patients. In order to manage patient information effectively in digital data format a web application is built. This web application registers users (professionals) that are responsible to register patients and includes a database to store patient’s information. Following this, custom genetic algorithm (GA) is developed considering constraints primarily patient category and the rest constraints such as patient in date and time, number of fractions, number of machine and also working days and working hour. Moreover, for the GA to be user friendly a desktop application with graphical user interface (GUI) is developed. The GUI supports the medical professionals to easily manipulate the GA parameters such as number of populations, crossover probability, and mutation probability and also change the dynamic resources or attributes like number of machines, number of patients treated per single machine and number of working days. As a result, the medical professional can schedule patients dynamically. In this thesis best GA performances (i.e., fitness value of 88% - 96.67% accuracy) are obtained for probability crossover (Pc) values between 60% - 80% and probability of mutation (Pm) between 20% - 40%. This means if the health professional sets the cross-over and mutation probability in these ranges, the scheduling will have better optimization, i.e. prioritize high-risk patients, minimize high risk patient waiting time, thus better care for patients. From the resuls, emergency patients are able to get early treatment than radical patients. Compared to traditional manual scheduling, where scheduling is done based on patients arrival date, GA based scheduling enables to prioritize higher risk patients.
Investigation of Photo-Induced Actuation Achieved Through Electro-Spun PLA/MWCNT Nanocomposite Fibers for Artificial Skeletal Muscle Fabrication
(Addis Ababa University, 2022) Rakeb Debalkew; Gyeong-Man Kim (Prof.)
Introduction: One of the major challenges in the medical industry is the area of technical orthopedics. Technical orthopedics is a branch of rehabilitation engineering that deals with the design of artificial limbs for people with physical disabilities. Plenty of materials and methods have been presented to the scientific community so far, but there still is a considerable gap in replacing the function of human skeletal muscle. This research project deals with the morphological and thermo-mechanical analysis of a PLA/MWCNTs nanocomposite fiber that can be applied as an actuator in the area of technical orthopedics, particularly for the fabrication of artificial muscles. Methodology: The characterization techniques output of the PLA/MWCNT nanocomposite fiber has been provided by Martine Luther University, Halle-Wittenberg, Germany. The morphological characterization outputs from SEM technique making use of image analysis software Image J and the mechanical characterization outputs from the Dynamic Mechanical Analysis (DMA) technique by theoretical approach and mathematical approach through MATLAB are analyzed to compare the behavior of actual human skeletal muscles with the proposed nanocomposite fiber. Results and Discussion: The output of the SEM characterization analysis clearly showed that there is Gaussian distribution of fiber diameter. Both manual and automated measurements were performed and both presented a normal distribution function with equivalent measurements of 6.7 μm and 7.7 μm mean respectively. The Gaussian distribution shows that the fabricated nanocomposite fiber has high reactivity with an external stimulus, in this case infrared radiation (IR). As for the DMA, the storage modulus showed a significant change from 327MPa to 371 MPa & the loss modulus increased from 62.5 MPa to 76.1 MPa up on the addition of MWCNT in the pure PLA. This shows improved mechanical (Elastic and Viscous) behavior of the nanocomposite fibers. The viscoelasticity of the fiber was also modeled by making use of the three element (Standard Linear Solid) Model and the temperature dependent modulus change is presented in this master thesis. Conclusion and recommendation: From the results acquired, the fabricated nanocomposite fibers showed properties that proved applicability for the fabrication of artificial human skeletal muscle. The fibers can also further be used for the detailed study and analysis of several forms of muscular malfunctions, such as, paralysis, cerebral palsy etc.
Human Activity Recognition using Machine Learning
(Addis Ababa University, 2023-08) Rediet Desalegn; Melkamu Hunegnaw (PhD)
Human activity recognition is a broad field of study concerned with identifying the specific movement or action of a person based on sensor data. It can improve people’s well-being by automatically assessing and summarizing their daily activities. In this thesis an automatic detection of nine different human activities from body sensor data was developed. The human activity data was collected from 30 participants between the age of 12 and 55. From those 17 participants were male and 13 participants were female. Participants completed nine activities (sitting, standing, walking, running, upstairs, downstairs, sit-up, jumping and cycling) while wearing an I-phone eight plus in the pocket. Records from 18 individuals was used for training, 6 participants data was used for validation and the remaining 6 records were used for testing. The human activity data is captured via phone’s integrated acceleration and gyroscope sensors. Hence a total of nine-dimensional data (triaxial accelerometer data (linear acceleration and body acceleration) and triaxial gyroscope data) was acquired using these two sensors. In all cases the sample rate used to capture the data was 50Hz which is in phone application. In this study both classical machine learning and deep learning methods were used. In classical techniques four machine learning classifiers (SVM, KNN, logistic regression and decision tree) have been used to categorize the data. For this to work 18 set of features were carefully computed. Three deep learning models (CNN, LSTM and hybrid CNN-LSTM) have been also trained in the accurate detection of human activity classification. The activity recognition of these algorithms has been objectively measured by using accuracy, f1 score, precision, recall and confusion matrix. Under this study accuracy of SVM 89%, KNN 81%, Logistic Regression 87%, Decision Tree 79%, CNN 92%, LSTM 91% and CNN-LSTM 94% are computed. The highest performance accuracy of 94% was achieved when classifying using hybrid CNN-LSTM model. This indicates that with a simplified approach, multiple human activities can be reasonably detected. By analyzing these activities, one can promote a healthy life style for the subjects.
Numerical Simulation of Sampling Capabilities of Spiral Trajectory and Echo-Planer Imaging in Magnetic Resonance Fingerprinting
(Addis Ababa University, 2023-01) Samuel Melke; Dawit Assefa (PhD)
Magnetic Resonance Fingerprinting (MRF) is a novel method proposed to solve the limitations of quantitative Magnetic Resonance Imaging (qMRI). One of the ways MRF accelerates data acquisition is by using various sampling mechanisms to undersample the k-space. In this thesis, the effectiveness of spiral sampling and accelerated cartesian sampling using multi-shot Echo-Planar Imaging (EPI) are compared by keeping the other steps in the MRF framework constant. Dictionary atoms were generated by using Bloch simulation. During the data acquisition, a realistic simulation framework based on the Bloch equation is built and implemented in a MATLAB platform. An inversion-recovery balanced steady-state free precession sequence was used in simulating the series images as well as the dictionary. The dot product of the simulated signal evolution and the dictionary of predicted signal evolutions is used for pattern matching. To check the efficacy of the methodology, cylindrical and brain numerical phantoms were used. The respective percentage errors in T1, T2 and off-resonance quantification were 2.6%, 2.3%, and 14% for spiral-MRF and 39%, 43%, and 124% for EPI-MRF. The results showed that spiral undersampling produces superior results in a close match with the ground truth compared to multi-shot EPI, showing the great promises of spiral trajectory to be used as an effective sampling tool in MRF. The MRF simulator developed in this thesis work effectively simulates the image acquisition process of an MRI machine and has consistently produced accurate results. The obtained results are generally comparable to those reported in other studies that utilized real scanners
Optimization of Bio-Paint Production Using Aspergillus Niger Under Liquid State Fermentation
(Addis Ababa University, 2022-10) Ejigayehu Desalegn; Solomon Kiros (PhD)
Synthetic dyes are hazardous to living organisms and the environment in general. This has progressively led to an increased demand for natural colorants. The fungus kingdom has been extensively researched as a source of bioactive chemicals with many industrial applications. However, in Ethiopia, such studies are lacking. In this work, the synthesis of natural pigments by Aspergillus niger spp. in a submerged fermentation system was investigated. The species were obtained from the biodiversity institute and grown on PDA agar for 5-7 days at a temperature of 30 °C. Even if many factors influence pigment production from microbes, in this study, only pH (3–8), maltose concentration (0.024–0.072 g), and fermentation time (7–28 days) were examined. For pigment production, cultivation of Aspergillus niger spp. was carried out in potato dextrose broth supplemented with a carbon source (maltose) for a day and pH adjusted in shaking conditions (centrifugal force of 150 rpm). Natural paint ingredients were used with fungal dyes to produce bio-paint. To establish optimal cultural conditions, Design Expert version 13 was used. To determine the optimal setting, a Box-Behnken design under response surface methodology was used. A second-order polynomial equation was fitted to the data using multiple regression analysis. A UV-Vis absorbance measurement was employed to quantify the pigment concentrations. Thin layer chromatography and FT-IR spectroscopy were employed to molecularly characterize the pigment. In this study, the maximum pigment concentration of 357.425 a.u. was obtained for the optimized values of pH, maltose concentration, and fermentation time of 7.678, 0.045 g, and 27.953 days, respectively. The bio-paint was found to have values of pH, density, viscosity and drying time of 6.8 + 0.57, 1.3272 + 0.04mg/ml, 55 mPas + 1.34 and 30 min + 1.63 min, respectively, and demonstrated comparable properties with synthetic paints. When applied to a wall, the bio-paint showed remarkable color and coating performance. Fungal pigments could, therefore, be utilized in expanding bio-colored industries in Ethiopia.
Classification of Fatty Liver Disease Using Deep Learning Technique
(Addis Ababa University, 2023-08) Hume Degebassa; Dawit Assefa (PhD); Million Molla (Mr.) Co-Advisor
Fatty liver disease (FLD) also termed steatosis liver disease can be classified into two broad spectrums: Alcoholic fatty liver disease (AFLD) and Non-alcoholic liver disease (NAFLD). AFLD is a type of liver disease associated with alcohol intake of 40 to 80 grams of ethanol/day for males and of 20 to 40 grams/day for females. NAFLD is a disorder characterized by excess accumulation of fat (in the form of triglycerides) with an amount >5% in the hepatocytes. The progression of the disease is manifested by ongoing inflammation and consequent steatosis grade: G1 (mild steatosis), G2 (moderate steatosis), and G3 (severe steatosis). If no early strategies are adopted, FLD can progress to steatohepatitis (SH) which is a risk factor for liver fibrosis, cirrhosis, and hepatocellular carcinoma (liver cancer). Cirrhosis is currently the 11th most common cause of death globally and liver cancer is the 16th leading cause of death. In Ethiopia, cirrhosis was the 7th leading cause of mortality accounting for 24 deaths per 100,000 populations in 2019. Hence, early diagnosis is important to avoid advanced stages of FLD. In order to diagnose and grade FLDs, Ultrasound imaging is the most well-known technique. The visual assessment of these Ultrasound images is not only time taking and labor-intensive task but it is also prone to inter-observer variability. That calls for the development of an automated system that overcomes subjectivity and inconsistency in the screening process. Different studies in the literature have proposed the classification of FLDs using deep learning and machine learning techniques. However, most of these techniques are limited to binary classification and don’t consider the severity level. The current study aims to develop an automatic fatty liver classification and grading scheme using deep learning techniques. A total of 550 liver ultrasound images were collected from Zenode repository and image pre-processing was employed on the acquired images before the images were used to train the deep learning model. The effect of important deep network hyper parameters including batch size, learning rate, regularizer and epoch was investigated. Different pre-trained deep learning networks were tested for their classification efficacy including VGGNet, MobileNetV2, ResNet and Xception based on useful performance matrices. Accordingly, the Xception model was found to outperform the rest for multiclass (normal, mild, moderate, and severe) fatty liver classification. It offered a precision of 96.25%, recall of 97.75%, F1 score of 96.75%, and overall accuracy of 96.36% showing the great promises of the model.
Improving Ultrasound Kidney Stone Detection Using Deep Learning
(Addis Ababa University, 2023-04) Bekalu Gedifew; Dawit Assefa (PhD)
Background: Currently a CT scan is preferred over Ultrasound images for Kidney stone diagnosis. The major problem regarding diagnosis of these stones under CT imaging mo-dality is that once these patients are diagnosed as positive; there is a high chance for the stone to be formed again in the patient’s lifetime after removal. As a result, use of the CT modality repeatedly exposes the patient for unwanted radiation exposure. Purpose of the Research: In ultrasound imaging, additive and multiplicative noises are taken as disadvantage for its imaging output. However, its ability to form real time imaging makes it preferable in many diagnosis procedures. The current study is aimed at developing an effective kidney stone detection scheme using a Convolutional Neural Network (CNN) by incorporating useful image pre-processing tools applied on Ultrasound images. Methods: The approach implemented in the proposed kidney stone detection scheme mainly involves two stages. The first stage employed using useful pre-processing steps applied on the Ultrasound images, which include image filtering, contrast enhancement and 2-D Directional Wavelet Transforms. The second stage is employed using multiclass se-mantic segmentation CNN models, which include Deep Lab V3 , U-net and Seg-net mod-els. In order to detect multiclass regions of ultrasound kidney stone image.The performance of the models was evaluated using useful quantitative matrices. Results and Conclusion: Results have shown that the Deep Lab V3 CNN model had greater performance than U-net and Seg-net CNN models tested in this study. The model was able to maintain a global accuracy and mean accuracy of 95.1% and 80.9% respec-tively showing its great promises in improving the detection of kidney stones based on Ultrasound images. Compared to performances reported in the literature by previous schol-ars who have developed different method of kidney stone detection algorithms, the pro-posed method has offered commendable results in terms of global accuracy and mean ac-curacy.
Design and Development of Steam Sterilizer
(Addis Ababa University, 2023-06) Yemane Tesfay; Dawit Assefa (PhD)
According to the 2010 WHO infectious disease report, hundreds of millions of patients around the world are affected by healthcare-associated infections which cause extended clinic stays, long-term disability, massive additional costs, and unnecessary deaths. This folds higher in low-income countries but it is a global burden. To minimize the burden of these infections, cleaning, disinfection, and sterilization techniques have been applied. For a sterilization technique, different steam sterilizers are available in the market ranging from the simplest electro-mechanical portable systems to the complex microcontroller-based double-door floor fixed types. However, those steam sterilizers can’t adjust environmental pressure for altitudinal changes that often fail to function properly under significant pressure variation. Countries like Ethiopia have great geographical diversity, known altitude ranges from the highest peak at Ras Dashen (4,620m) down to the Dallol Depression (-148m), which highly affect the performance of steam sterilizers. The main intent of this thesis study is to design a microcontroller-based smart steam sterilizer that adjusts itself to variable altitudes since steam sterilizers that work in all altitudes are vital for the healthcare system. The sterilizer also comes with a water electrical conductivity sensor to completely avoid the use of hard water integrated with a water level regulator to protect the heater when water runs out inside the chamber. Interfacing between the proposed smart pressure compensator, conductivity sensor with water level regulator of steam sterilizers is worked out in this study. The developed control system senses the environmental condition using BMP280 sensor to adjust itself based on the altitudinal changes as well as water electrical conductivity, water level, pressure, and temperature sensors. A formula is developed based on the altitude pressure compensation to solve the problem of altitudinal effect and for water quality and quantity measurement guided by a programmable microcontroller. The control system can adjust the pressure based on the altitudinal variations. The system offers better efficiency, precision and comes with great ease of use. The designed smart steam sterilizer is as effective as existing modern sterilizers with an added feature of blocking use of hard water. Meanwhile, further research might be needed to make the device more efficient, lower cost, durable, and safer.
A Trilingual Android Application with Automatic Malaria Detection from Microscopic Images of Red Blood Cells
(Addis Ababa University, 2023-12) Berihun Nigussa; Dawit Assefa (PhD)
Malaria, which is a mosquito-borne blood disease caused by Plasmodium parasites, is one of the virulent infectious diseases affecting human beings and other animals since antiquity. Even though there were promising progresses in the reduction of malaria morbidity and mortality in the past two decades before the outbreak of COVID-19, the latest two reports of theWorld Health Organization (WHO) statistics indicate that malaria has been overlooked due to the COVID pandemics. Malaria is still prevalent specifically in low resource setting areas such as the sub-Saharan African countries, including Ethiopia. WHO reported that there were 229 million new cases of malaria and 409,000 deaths globally in 2019, alone. Whereas in the year 2021, the morbidity and mortality was reported to rise up to 247 million and 619,000, respectively. Timely diagnosis and treatment as well as good awareness about the disease play a major role to combat malaria. In the current project work, it was intended to design and develop a multi-lingual Android App that offers useful information about the malaria disease and is capable of automatically detecting malaria infected red blood cells (RBCs) from color microscopic images based on a deep learning approach. The Convolutional Neural Network (CNN) based deep learning model was trained, validated and tested on a publicly available dataset composed of microscopic images of RBCs taken from individuals with confirmed malaria infection as well as normal control groups. Experimental results generated from the deep learning model showed that the detection capability of the model achieved 100% training accuracy, 96% validation accuracy and 96% testing accuracy. The developed App avails useful information about malaria disease in general and tips users with fundamental information regarding its prevention and transmission mechanisms acting as an m-health system.
Design and Evaluation of Simulated PEM Scanner for EarlyStage Breast Lesion Detection
(Addis Ababa University, 2022-06) Bethel, Haile; Dawit, Assefa; Samuel, Taddesse (Mr.) Co-Advisor
Breast cancer is the most commonly diagnosed cancer in females worldwide, contributing around 11.7 % (~2.3 million people) of new cancer cases in 2020 only, with a death rate of 6.9%. Early diagnosis and check-up are essential for effective treatment and reduction of incidences and mortality rates. Different modalities exist to diagnose breast cancers. Positron emission tomography (PET) is the one imaging tool in nuclear medicine providing physiological information about the breast by quantifying the metabolic activities of the cells in the breast tissues. Both whole-body (WB) PET scanners and organ-specific PET scanners are available in the market. But current trends show that there is a complete shift towards dedicated, organ specific PETs. One of those is Positron Emission Mammography (PEM), believed to be better than mammography and other imaging modalities to detect small breast lesions. Semiconductor-based PEM detectors are simulated with good spatial resolution but are expensive. Scintillator-based PET and PEM detectors can provide quit good sensitivity and are cost effective. On the contrary, these detectors cannot detect small breast lesions due to their poor spatial resolution. This requires development of detectors that give rise to better spatial resolution. In the current thesis work, a high-performance PEM scanner is simulated using TOC (Transparent Optical Ceramic) scintillators of 1 x 1 x 10mm 3 crystals with the aim to improve the spatial resolution, sensitivity as well as scattering fraction. Those TOC scintillators are LHO: cerium doped lutetium hafnate (Lu 2 Hf 2 O 7 : Ce), BHO: cerium doped barium hafnate (BaHfO 3 : Ce) and SHO: cerium doped strontium hafnate (SrHfo 3 : Ce). The design was based on the GATE (Geant4 Application for Emission Tomography) simulation software. Its performance was tested and evaluated by following the NEMA (National Electrical Manufacturers Association) NU 4-2008 standards. The complete scanner has 39 heads and 10 x 30 x 59 modules in the detector. Based on the type of scintillator used, the designed scanner provided a spatial resolution between 1.0 and 1.1 mm FWHM (Full Width at Half Maximum) in the axial direction, 7.24% to 9.11% system sensitivity and 11.01% to 11.19% scatter fraction. The design offered good uniformity as well as image quality. The computed spatial resolution, sensitivity and scatter fraction values are superior to those already reported in the literature.
Trinion Based WBC Segmentation Using Texture to Detect Acute Lymphoblastic Leukemia
(Addis Ababa University, 2021-10) Amin, Hussien; Dawit, Assefa (PhD)
Many diseases are detected based on examination of microscopic images of blood samples. Changes in the blood condition show the development of diseases in an individual. One type of disease caused by change of blood condition is Leukemia. Leukemia can cause early death when it is not treated on time. In Ethiopia, Leukemia accounts to about 35.5% of hematological admissions. The death rate in Ethiopia due to Leukemia is different with time and region. The average death rate is increasing from time to time and shows variation between country side and urban population. Reports from the World Health Organization (WHO) show that death rate due to Leukemia in Ethiopia has reached 5.56% and ranked 18 th highest in the world. Leukemia originates in the bone marrow, a thin material inside the bone. Leukemia is detected by analyzing white blood cells (WBCs also called Leucocytes), one of the constituents of blood along with red blood cells (RBC or Erythrocytes), platelets and blood plasma. WBCs have five different types (Lymphocytes, Myelocytes, Neutrophils, Basophils and Eosinophils) and among these Lymphocytes and Myelocytes are the ones that could start to change in the bone marrow and get infected and become Leukemic or infected cells. These Leukemia cells have strange properties compared to the normal cells in that their growth is abnormal and they survive much longer than the normal cells. They also interrupt functions of the normal cells. Through time, the normal cells perish while leukemia cells still survive. Old leukemia cells last for a longer time and production of new leukemia continue in an abnormal way. Traditionally, Leukemia detection is carried out manually based on visual examination of microscopic images of blood samples. This is lengthy and time taking process which depends on the skills and experiences of the observer which makes the process subjective. In this regard, computer based automated schemes play their great role and several efforts have been made in the literature to develop such schemes. In the current study, a novel mathematical technique for Leukemia detection based on holistic analysis of color microscopic images of blood samples is proposed. The approach utilizes a holistic representation of microscopic blood images in the three (Trinion) space and applies trinion based Fourier transform implemented in the L*a*b color space to extract useful higher order features to segment normal and infected WBCs and classify them. The technique has been applied in analyzing microscopic images acquired from standard ALL-IDB database. Classification of normal and Leukemic WBCs was performed based of Artificial Neural Network (ANN) which resulted in 95.7% sensitivity, 100% specificity and 97.6% accuracy.
Application of Virtual Reality for Teaching in the Dissection Room
(Addis Ababa University, 2021-09) Alemseged, Getachew; Dawit, Assefa (PhD)
Virtual reality (VR) is a computer graphics technology which can be used to develop applications to support teaching human anatomy. Using VR, virtual 3D models of various parts of the human body can be developed and interacted with. The present thesis work addresses the problem that medical students often encounter in their practice in wet specimens with the absence of cadavers, using kidney as an example. The main objective of the thesis is to reconstruct highly detailed 3D virtual model of the kidney based on small interval cross- sectional images (CT-DICOM dataset), suitable for undergraduate students teaching virtual dissection through animation using kidney as an example, to supplement the wet specimen dissection procedure.. A 3D slicer software is used to develop 3D model of the kidney from the imported CT-DICOM data. Output of the 3D slicer is high poly object which is exported in an STL file format to the Transmutr software for simplifying the number of faces of the model by preserving the overall shapes of the original object which is used by Google sketch up & SimLab Composer software. Google Sketch Up is used for dissecting the kidney while SimLab Composer is used for rendering the 3D image and finally used for interactive VR viewing purpose. The final results of the developed 3D model of the kidney are used for interactive virtual reality viewing using desktop computer and virtually dissecting the kidney using Google sketch Up for visualization of the internal structure of the developed model. The developed 3D model allows students to perform detaching and interactively viewing a dissected kidney sequentially without the need of a supervisor. It can be also used inside the dissecting room to practice virtual kidney dissection by other medical practitioners.
Genetic Algorithm based Optimized Radiotherapy Patient Scheduling
(Addis Ababa University, 2022-02) Metadel, Hailu; Dawit, Assefa
Radiotherapy is the major means to treat cancer patients. Radiotherapy comprises two phases: pretreatment and treatment on radiation machines. This thesis work focuses on the treatment phase. Treatment consists of multiple, almost daily irradiation appointments, followed by optional imaging and control assignments. The scheduling of radiotherapy appointments is a complex problem due to various medical and scheduling constraints, such as patient category, machine availability, waiting time targets and also due to the size of the problem (i.e., number of machines, facilities and patients). The objective of this thesis is to minimize waiting time and maximize device utilization in) patient’s appointment scheduling. Thus, this thesis presents an optimization algorithm for scheduling of radiotherapy treatments for categorized cancer patients. In order to manage patient information effectively in digital data format a web application is built. This web application registers users (professionals) that are responsible to register patients and includes a database to store patient’s information. Following this, custom genetic algorithm (GA) is developed considering constraints primarily patient category and the rest constraints such as patient in date and time, number of fractions, number of machine and also working days and working hour. Moreover, for the GA to be user friendly a desktop application with graphical user interface (GUI) is developed. The GUI supports the medical professionals to easily manipulate the GA parameters such as number of populations, crossover probability, and mutation probability and also change the dynamic resources or attributes like number of machines, number of patients treated per single machine and number of working days. As a result, the medical professional can schedule patients dynamically. In this thesis best GA performances (i.e., fitness value of 88% - 96.67% accuracy) are obtained for probability crossover (Pc) values between 60% - 80% and probability of mutation (Pm) between 20% - 40%. This means if the health professional sets the cross-over and mutation probability in these ranges, the scheduling will have better optimization, i.e. prioritize high-risk patients, minimize high risk patient waiting time, thus better care for patients. From the resuls, emergency patients are able to get early treatment than radical patients. Compared to traditional manual scheduling, where scheduling is done based on patients arrival date, GA based scheduling enables to prioritize higher risk patients.
Toward A Simulation Framework For Magnetic Resonance Fingerprinting Using Spiral Underdamping
(Addis Ababa University, 2020-12) Muhammed, Ezriku; Dawit, Assefa (PhD)
Magnetic Resonance (MR) is a powerful and multipurpose measurement technique to look inside the human body using magnetic properties of the body non-invasively. Quantification of tissue properties including the relaxation parameters has long been a goal of magnetic resonance imaging (MRI), to provide a basis for diagnosis, longitudinal study and, inter-patient comparability. However, prolonged acquisition times have hindered the usage of quantification for clinical applications. Magnetic Resonance Fingerprinting (MRF) was introduced as a promising technique for simultaneous and fast quantification of multiple tissue parameters through a new approach to data acquisition and postprocessing. However, the development and optimization of MRF process is time-consuming, expensive which requisite repetitive experiment and often needs a synthetic phantom or human subject to test it on the real scanner and detect the results. This work aimed to develop an implementation of a Simulation Framework for MRF based on spiral under-sampling. Simulation framework for magnetic resonance fingerprinting (MRF) along with evaluating the effect of noise on the parameter map is significantly studied by using custom generated phantom and phantom generated from the brain web of a simulated database. The undersampling capability of MRF is significantly studied and evaluated by comparing the parameter map that is generated at each undersampling level. Significant undersampling is used to make the MRF time-efficient with several undersampling folds and undersampling with the factor of 24-fold is substantially evaluated and results in acceptable tissue quantification result irrespective of the undersampling artifacts. A comparison between the proposed MRF simulation and the existing simulation used for MRF was done and the proposed methods showed superior results to simulate the effect of gradient on the reconstructed image and at the quantified parameter map. The effect of noise on the parameter map is evaluated by adding different levels of noise on the simulated kspace signal. At a given undersampling level, as the level of noise increases, its effect on the parameter maps gets more and more pronounced specifically in the off-resonance map.

Browse

Recent Submissions