Browsing by Author "Daamouche, Abdelhamid (Supervisor)"

Now showing 1 - 9 of 9

Atrial fibrillation analysis by deep learning.
(Université M’hamed Bougara de Boumerdes : Institut de Genie Electrique et Electronique, 2024) Agli, Wafa; Daamouche, Abdelhamid (Supervisor)
Atrial fibrillation (AF), an increasingly prevalent cardia carrhythmia, is a major contributor to stroke, heart failure, and premature mortality. Traditional manual screening for AF using electrocardiography (ECG) is not only time-consuming but also susceptible to human error, underscoring the urgent need for automated diagnostic tools. This study addresses this challenge by developing advanced computer-aided diagnostic methods leveraging deep learning for the automatic detection of AF. We introduce innovative one-dimensional (1D) and two-dimensional (2D) convolutional neural network (CNN) models specifically designed for the precise classification of ECG signals into normal or atrial fibrillation categories. Our methodology includes a meticulous preprocessing phase where each ECG record is filtere dan dpeak sare accurately detected using the XQRS algorithm. The signals are then segmented into beats with an 80-sample window, which serve as critical features for subsequent classification. The extracted features are fed into our CNN architectures for classification. The models are trained and evaluated using the MIT-BIH Atrial Fibrillation Database, and their generalization capability is further validated with unseen data from the PhysioNet/Computing in Cardiology Challenge 2017 database, following an inter-subject approach. To enhance the robustness of our models, we employ data augmentation techniques. Our comprehensive evaluation demonstrates that the 1D-CNN model achieves a remarkable total accuracy of 95% and an F1 score of 96.81%, while the 2D-CNN model attains an exceptional accuracy and F1 score of 99.84%. These results underscore the efficacy of our approach in accurately classifying ECG signals and highlight the potential of our models for real-world clinical applications, offering a substantial improvement in AF screening processes.
Cell nuclei segmentation in histopathology images based on deepneural networks
(2022) Merah, Halima; Zemouri, Nassima; Daamouche, Abdelhamid (Supervisor)
Despite the significant progress in understanding cancer’s biological basis, it continues to confound patients, researchers, and physicians as a self-sustaining and adaptive disease that interacts dynamically with its environment. Analysis of stained tumor sections has a staggering importance in cancer diagnosis and prognosis, which is mainly carried out manually by pathologists. Because most of the human body's billions of cells have a nucleus full of DNA, the genetic code that programs each cell, most analyses begin with identifying the cell's nuclei. Researchers can better grasp the underlying process by identifying the nuclei. They can measure how different samples react to a certain drug. However, huge volumes of medical images make manual analysis challenging, time consuming, and a tedious task. Therefore, other techniques are necessary to automatically analyze large amounts of these complex image data in order to draw biological conclusions from them and to study cellular and tissular phenotypes at a large scale. The automatic segmentation of cell nuclei from this type of image data is one of the bottlenecks for such techniques. We present a fully automated workflow to segment nuclei from histopathology images by using deep neural networks trained from a set of semi automatically annotated image data. In our work, we have used the PSB 2015 crowd-sourced nuclei dataset. We have built three models using three different architectures: U-Net, U-Net++, and a modified version of U-Net architecture.
Chest medical image classification using deep network.
(Université M’hamed Bougara de Boumerdes : Institut de Genie Electrique et Electronique, 2023) Tis, Mohammed Amine; Daamouche, Abdelhamid (Supervisor)
Human lung which is among the most important parts in human body is facing mortal diseases especially after the COVID-19 pandemic. The scientific world is rapidly developing the health-care field to face these disorders and save millions of lives all around the world. The primary objective was to find aprecise an defficient strate gyf ort heaccura tea ndear lydetecti onand classificatio no flun gdiseases .T oachiev ethi sgoal ,w euse dth epowe ro ftwo essential medical imaging techniques: computerized tomography (CT-scan) and X-ray imaging. Additionally, we employed three deep learning models: Inception-v3, ResNet, and DenseNet, coupled with two distinct classification; binary classificatio nan dmulti-clas sclassificatio n.O urresear chjourn eystarted with binary classification ,focusin go ndistinguishin gbetwee nCOVID-1 9and non COVID-19, using both CT-scan and X-ray datasets in total of 17,599, all three models delivered outstanding results, with the highest accuracy reaching an impressive accuracy of 96%, achieved by DenseNet using CT-scan images. These results underscore the potential of deep learning in helping healthcare professionals with highly accurate disease classification .Shiftin gt oth emulti- class classificatio ndictate db yth enee dfo r amor ecomprehensiv ean drealistic approach to diagnosing and identifying a wide range of medical conditions in clinical practice and research. The new class added to COVID-19, non COVID-19 is: Community-acquired pneumonia (CAP), in total of 17,104 CT- scan images,and using the same models we challenged the system using different splitting data ratios. Through a series of experiments and evaluations, our system achieves an overall accuracy of 98% in classifying chest images across multiple categories, using DenseNet model and the 80:10:10 splitting ratio. The results showcase the significan tpotentia lo fdee plearnin gi nassisting healthcare.
Classification of ECG Signals Using Deep Learning
(Université M’Hamed bougara : Institute de Ginie électric et électronic, 2023) Zanaz, Serine; Kermane, Imane; Daamouche, Abdelhamid (Supervisor)
Accurate classifi cation of electrocardiogram (ECG) signals is crucial for diagnosing cardiac conditions. In this project, our objective was to classify ECG beats into disease classes using deep learning techniques. We leveraged two primary datasets: the MIT-BIH dataset from PhysioNet and the INCART 12-lead Arrhythmia Database from St. Petersburg, providing a comprehensive basis for our classifi cation models. Our methodology involved a hybrid model combining 1D and 2D convolutional neural networks (CNNs). We applied a 1D CNN architecture to process ECG signals directly and transformed ECG beats into images for a 2D CNN architecture. By incorporating both approaches, we captured temporal and spatial information in the ECG signals. Data augmentation techniques were employed to address imbalanced data distribution and improve model performance.
Deep learning-based mobile application for plant disease diagnosis.
(2022) Beggar, Ikram; Daamouche, Abdelhamid (Supervisor)
Plants are susceptible to a variety of diseases in farming. The impact of sudden climatic change harms their growth causing dangerous viruses and pests. Plant diseases are one of the most serious problems confronting agriculture across the world and harming the health, economy, and livelihood of the human population. The majority of traditional plant pests' diagnosis methods rely on human visual observation and inspection. However, this approach is time-consuming and requires strong agricultural skills and significant human efforts. Recent breakthroughs in computer vision and Deep Learning provide a potential pathway for developing a plant disease diagnosis system that will be able to detect plant diseases in different geographical regions with fewer human interventions. This dissertation presents a Deep Learning powered mobile-based system to automate the early identification of plant diseases. The developed system uses Convolutional Neural Networks to classify plant leaves into 38 classes with 14 types of plants. To train and test our models, we used two datasets from Kaggle containing about 87000 and 20803 images of healthy and diseased plant leaves respectively. After the training, we assessed our models using some classification evaluation metrics such as the accuracy and f1-score and found that they were able to correctly classify most of the images in the test set. Finally, to increase the usability of our algorithm, we developed a smartphone application that runs on both Android and iOS operating systems with a simple user interface using the Flutter framework. This application allows farmers to capture pictures of their infected plant leaves or import them from their phone library and then it displays the disease category with the plant name and the accuracy of the prediction. This approach is supposed to provide farmers with a better opportunity to maintain their crops' health and minimize the use the incorrect fertilizers that might damage the plants and the environment.
Diabetic retinopathy grading using deep neural networks
(Université M’hamed Bougara de Boumerdes : Institut de Genie Electrique et Electronique, 2023) Temmam, Amira; Ahmed Gaid, Chaima; Daamouche, Abdelhamid (Supervisor)
Diabetic Retinopathy (DR) is a chronic disease and the leading cause of blindness and visual impairment among diabetic patients making early detection and classificatio nof DR crucial for effectiv etreatment .Thi sprojec tutilize sstate-of-the-ar tconvolutiona lneu- ral networks and transfer learning techniques to analyze retinal images and classify the severity of the disease on a scale of 0 to 4 (ranging from healthy to proliferative). We employ architectures such as EfficientNetB 1,InceptionV 3,Xceptio n,a ndMobileNet V2to achieve accurate classification .Throug h aserie so fexperiment san devaluations ,ou rsys- tem achieves an overall accuracy of 80.0% in classifying DR images. The results showcase the significan tpotentia lo fdee plearnin gi nassistin ghealthcar eprofessional swit hearly diagnosis and treatment planning, thereby improving patient well-being.
Dimensionality Reduction & Classification of hyperspectral images using Deep Learning
(2022) Hemissi, Mahfoud; Daamouche, Abdelhamid (Supervisor)
Hyperspectral image classification is a widely used technique for the analysis of remotely sensed images, due to the rich spatial and spectral information the hyperspectral images provide. Therefore, they are used in many real world applications. However, hyperspectral images classification can be a challenging task. On one hand, the large number of spectral bands yields to a complexity of the system. On the other hand, the task may result in a poor performance due to the high inter-class similarity and high intra-class variability. Towards this direction, we proposed deep learning methods to extract the most important features from the hyperspectral images. We focused on the convolutional neural networks because they have the ability to learn high-level spatial and spectral features; they process the input data to generate the classification outcomes. The experiments have been performed on four datasets, namely: Indian Pines, Salinas Scene, Pavia University, and Kennedy Space Center. The results obtained using the CNN algorithms were considerably high, and were predominant to other machine learning and deep learning algorithms.
Multi-labeled chest X-Ray images classification using transfer learning
(Université M’hamed Bougara de Boumerdes : Institut de Genie Electrique et Electronique, 2023) Messouci, Bouchra; Lallem, Manel; Daamouche, Abdelhamid (Supervisor)
Chest diseases are a prevalent global health issue affecting millions of people worldwide. They can vary in severity with some conditions being relatively mild and others posing serious health risks, if not diagnosed and treated promptly. One of the most common imaging techniques used to diagnose chest pathologies is x-ray .It offers a non-invasive, quick, and relatively cheap mean to gain insights into the internal structures of the chest. However, the interpretation of chest x-ray images can be challenging. Factors such as overlapping structures, variations in image quality, the presence of more than one abnormality in one image, and the need for different viewpoints can complicate the interpretation process. Radiologists and healthcare professionals require specialized training and a high level of expertise to be able to accurately analyze these images and identify potential chest diseases. In this study, we investigate the effectiveness of deep learning methods in detecting pathologies present in chest radiographs. Specifically, we focus on the application of convolutional neural networks for classifying different types of pathologies. Convolutional neural networks have gained popularity due to their capability to learn meaningful image representations at various levels. Our research explores the potential of using networks trained on a non-medical dataset for the multi- labeled classificatio no fvariou sches tpathologies .T oevaluat eth eperformanc eo four algorithms, we used a subset of the CheXpert public data-set available on Kaggle. We also investigated the effect of class balancing using various techniques on the overall performance. Among the different approaches we tested, EfficientNet-B2trained on the balanced data-set yielded the best results. For the various types of pathologies, we achieved an area under the curve (AUC) ranging from 0.850 to 0.876 These results demonstrate the feasibility of utilizing transfer learning approaches to detect pathology in chest X-rays.
Nuclei segmentation using deep neural networks
(2021) Benyoussef, Ahmed; Aissaoui, Marwan Abdel Illah; Daamouche, Abdelhamid (Supervisor)
One of the most important tools incancer diagnosis, prognosis, and grading is the analysis and interpretation of stained tumor sections, which is mostly done manually by pathologists. With the advent of digital pathology, that provides us with challenging opportunity to automatically analyze huge amounts of these complex image data in order to derive biological conclusions from them in order to study cellular phenotypes on a wide scale. The automatic segmentation of cell nuclei from this type of image data is one of the bottlenecks for such techniques. Cell nucleissegmentation is essential for a variety of bioimaging tasks, including cell counting and tracking, cell morphology characterization, and molecular expression quantification. Accurate automatic nuclei segmentation is of special interest in high-throughput applications of microscopic images of cells or tissues. In the image processing world, cell nuclei segmentation is an open challenge and a hot topic of research. We used a fully automated approach for segmenting nuclei from histopathology image data using deep neural network trained on a set of manually annotated images from scratch. The dataset that we used in our work was provided by the Department of Biomedical Engineering, Case Western Reserve University, USA [48]. We built our deep neural network using a modified version of U-Net architecture. To evaluate our model, we used three different metrics which are the Pixel Accuracy (PA), Intersection over Union (IoU), and the dice-coefficient. The results obtained are as follow: 0.98 for PA, 0.57 for IoU, and 0.38 for dice-coefficient.