Browsing by Author "HAZMOUNE , Samira"

Now showing 1 - 3 of 3
  • Thumbnail Image
    Item
    A Transformer-based ensemble frame work for Early diagnosis of genetic syndromes using facial feature analysis
    (Faculty of Sciences, 2025) BELHACINI , Amila; LEMOUCHI , Ala Rahma; HAZMOUNE , Samira
    This thesis ,titled "A Transformer-Based Ensemble Frame work for Early Diagnosis of Genetic syndromes Using Facial Feature Analysis", presents an advanced deep learning approach for early detection of pediatric genetic syndromes using facial images. The motivation stems from the need to assist clinicians with faster and more accurate diagnosis , as facial traits often contain key indicators of genetic abnormalities .To address the limitations of traditional convolutional models, this work integrates Transformer architectures capable of capturing global dependencies and fine-grained facial patterns that are essential for distinguishing between syndromic and non-syndromic faces. The proposed system employs three state-of-the-artTransformer-models- based-Vi-sion Transformer (ViT),Data-efficient Image Transformer(DeiT),and Swin Transformer (Swin-T) —which are combined through ensemble learning using Random Forest,XG-Boost , and Logistic Regression meta-classifiers. This hybrid strategy enhances generalization and robustness across diverse facial representations. Experimental results on a custom-built dataset demonstrate a significant improvement in accuracy, achieving 86%, validating the effectiveness of the ensemble Transformer framework for automated pediatric syndrome diagnosis.
  • Thumbnail Image
    Item
    Transformers-based Ensemble Methods for Medical Imaging: A Theoretical and Experimental Study
    (Faculty of Sciences, 2024) ZIANE , Selma; HAZMOUNE , Samira
    Medical image classification accuracy significantly aids doctors in diagnosing diseases and planning treatments. Although Transformers, a cutting-edge technology in artificial intelligence, perform exceptionally well in various image classification tasks, their ability to capture the subtle details within medical images can be limited. To address this challenge, we propose, in this thesis, a novel ensemble method that leverages Transformers to improve medical image classification. We employ eight diverse medical imaging datasets and implement eight different Transformer-based ensemble methods on each dataset, covering several medical imaging areas such as magnetic resonance, computed tomography, magnetic resonance, dermatoscopic, chest X-ray, screening Mammography of medical imaging classification. A deep experimental study of the proposed approach is conducted to evaluate its effectiveness in the medical image classification domain. Furthermore, an ablation study is performed to identify the optimal combination of base models for each ensemble method across different datasets. Our experiments encompass datasets of varying sizes, acknowledging the ongoing challenges of limited data availability in medical imaging. Despite this limitation, our ensemble method approach consistently outperforms state-of the-art methods across multiple datasets. This demonstrates its effectiveness in alleviating limitations associated with data size and diversity. These findings highlight the potential of Transformers-based ensemble methods to revolutionize medical image classification. This paves the way for improved diagnostic accuracy and treatment decision-making in clinical settings.
  • Thumbnail Image
    Item
    Un Modèle d’Apprentissage Profond pour la Reconnaissance des Caractères Manuscrits
    (Faculté des Sciences, 2022) SAADI , Selma; HAZMOUNE , Samira;
    Ce mémoire s’inscrit dans le cadre de la reconnaissance automatique de l’écriture manuscrite, qui a suscité beaucoup d'intérêt de la part des chercheurs au cours des dernières décennies, en raison de son omniprésence dans les domaines où les humains interagissent, communiquent et effectuent des transactions. Le Deep Learning (Apprentissage Profond) et plus particulièrement, le réseau de neurones convolutif CNN (Convolutional Neural Network) a montré beaucoup de succès dans le domaine de reconnaissance de l’écriture manuscrite. Le succès est largement lié à son architecture profonde qui comporte plusieurs couches cachées permettant l’extraction automatique des caractéristiques. La modélisation des couches CNN ainsi que les hyper paramètres ont un effet profond sur la précision du modèle de reconnaissance. Dans ce mémoire, nous présentons un modèle de Deep Learning basé sur les CNN pour la reconnaissance des caractères manuscrits. Afin d’améliorer la précision du modèle, une série d’expérimentations sur l’ensemble de données EMNIST a été réalisée pour déterminer les hyper paramètres optimaux de l'architecture CNN. Le réseau profond CNN a été entrainé et évalué sur les sous-ensembles de données EMNIST Letters, EMNIST Balanced et EMNIST MNIST. La précision de la classification pour les trois ensembles de données était de 94,28 %, 88,93 %, et 99,32 % respectivement. Les résultats comparatifs avec certains algorithmes d’apprentissage et quelques travaux précédents sur la même base de données sont très encourageants.