Browsing by Author "Gaceb, Djamel"

Now showing 1 - 13 of 13

A new fast DBSCAN using dual-space analysis and colour integral volume for document image segmentation
(Inderscience Publishers, 2025) Kezzoula, Zakia; Gaceb, Djamel
The segmentation of the colour document images is an essential step allowing facilitating and improving the stages of characterisation and interpretation of the information contained in these documents. Recent systems of automatic processing and recognition of document images, which use separation of colorimetric layers, are more efficient compared to conventional systems, only based on binary or grey levels images. This task requires non-supervised pixel segmentation or clustering techniques to separate the document image to a variable and unknown number of colour layers. The methods based on density are widely used in this context at pixel level, such as the DBSCAN method and its different variants, very robust to the noise and more adapted to the degradations present on document images, but who suffer from a great complexity. In this context, we propose a new faster DBSCAN variant using the volume integral in colorimetric space for the first time to significantly reduce calculation time. The combination of the two spaces, Cartesian and colorimetric has also accelerated the method and improved its performance on document images with different challenges. The results obtained show the effectiveness of the proposed approach, which was marked by significant improvement in the quality of segmentation and reduction in computed time
Application of Deep Transfer Learning in Medical Imaging for Thyroid Lesion Diagnostic Assistance
(Institute of Electrical and Electronics Engineers, 2024) Chaouchi, Lynda; Gaceb, Djamel; Touazi, Fayçal; Djani, Djouher; Yakoub, Assia
This academic work evaluates and compares the performance of various deep convolutional neural network (DCNN) architectures in classifying thyroid nodules into two categories, malignant and benign, using ultrasound images. The dataset comprises 269 cases of benign lesions and 526 cases of malignant lesions. Given the limited dataset size, we employ a progressive learning approach with three established CNN models: VGG-16, ResNet-50, and EfficientNet. Initially pretrained on ImageNet, these models undergo further fine-tuning using a radiographic image dataset related to a different medical condition but similar to our domain. Different levels and fine-tuning strategies are applied to these models. A supervised softmax classifier is used for classifying lesions as malignant or benign, with the exception of the VGG-16 model. For the VGG-16 model, two additional classifiers, Support Vector Machine (SVM) and Random Forest (RF), are evaluated. The results obtained demonstrate the possibility of easily transitioning from the classification of one disease to another, even with a limited number of images, by leveraging the knowledge already acquired from another extensive database.
Blood cells image segmentation and counting using deep transfer learning
(IEEE, 2023) Gharbi, Aghiles; Neggazi, Mohamed Lamine; Touazi, Faycal; Gaceb, Djamel; Yagoubi, Mohamed Riad
In this paper, we present a two-step automatic blood cell counting approach for accurately and efficiently determining the complete blood count (CBC). The approach involves using two convolutional neural networks (CNNs) for the segmentation of red blood cells, white blood cells, and platelets, and then applying three different algorithms (Watershed, Connected Component Labeling, and Circle Hough Transform) to count the cells present in the masks produced by the CNNs. We also introduce a loss function for the Circle Hough Transform algorithm to further improve its accuracy. Our approach shows good results compared to other methods in the literature and has the potential to significantly reduce the time and effort required for manual blood cell counting
Comparative Evaluation of StyleGAN3-Based Augmentation Strategies for Enhanced Medical Image Classification
(CEUR-WS, 2025) Touazi, Faycal; Gaceb, Djamel; Tadrist, Amira; Bakiri, Sara
Deep learning models for medical image classification face significant challenges due to class imbalance and the limited availability of annotated datasets, particularly for rare diseases. Traditional data augmentation techniques, such as rotation, translation, etc., often fail to provide sufficient diversity to perform a good classification for minor classes. To address this issue, various strategies have been explored, including oversampling, undersampling, cost-sensitive learning, and synthetic data generation using generative adversarial networks (GANs). In this study, we evaluate the impact of using a generative AI based approaches and demonstrate that the most effective strategy is to combine synthetic augmentation with traditional methods. Specifically, we employ StyleGAN3 to generate high-fidelity synthetic images that, when integrated with traditional data-augmentation techniques, may improve the performance of deep learning models on medical image classification. We validate our method on datasets, including COVID-19 chest X-rays and HAM10000. Experimental results show that this hybrid approach leads to an improvement in classification accuracy, particularly for minority classes, surpassing standalone augmentation strategies. Our findings highlight the potential of AI-driven synthetic data generation as a complementary solution to traditional augmentation, offering a more balanced and diverse dataset for medical image analysis.
A Fast High Precision Skew Angle Estimation of Digitized Documents
(Senantic scholar, 2020) Chettat, Merouane; Gaceb, Djamel; Belhadi, Soumia
In this paper, we treated the problem of automatic skew angle estimation of scanned documents. The skew of document occurs very often, due to incorrect positioning of the documents or a manipulation error during scanning. This has negative consequences on the steps of automatic analysis and recognition of text. It is therefore essential to verify, before proceeding to these steps, the presence of skew on the document to be processed and to correct it. The difficulty of this verification is associated to the presence of graphic zones, sometimes dominant, that have a considerable impact on the accuracy of the text skew angle estimation. We also noted the importance of preprocessing to improve the accuracy and the calculation cost of skew estimation approaches. These two elements have been taken into consideration in our design and development of a new approach of skew angle estimation and correction. Our approach is based on local binarization followed by horizontal smoothing by the Run Length Smoothing Algorithm (RLSA) method, detection of horizontal contours and the Hierarchical Hough Transform (HHT). The algorithms involved in our approach have been chosen to guarantee a skew estimation: accurate, fast and robust, especially to graphic dominance and real time application. The experimental tests show the effectiveness of our approach on a representative database of the Document Image Skew Estimation Contest (DISEC) contest International Conference on Document Analysis and Recognition (ICDAR)
New CNN stacking model for classification of medical imaging modalities and anatomical organs on medical images
(CEUR Workshop Proceedings, 2023) Khaled, Mamar; Gaceb, Djamel; Touazi, Fayçal; Aouchiche, Chakib Ammar; Bellouche, Youcef; Titoun, Ayoub
Decision making in medical diagnosis is tedious and very rigorous task, hence the requirement to use more advanced and intelligent medical imaging diagnostic support systems. The automation of the recognition of medical imaging modalities and human anatomical organs gives these systems the possibility of processing, in an automatic and adapted manner, different types of images in consideration of different medical imaging modalities. It also offers better support to clinicians and patients allowing them to access to more effective image analysis and diagnostic tools. In this context, three deep learning approaches were developed and tested on six different CNN models (VGG16, VGG19, ResNet-50, Xcpetion, Inception and NASNet). Two deep transfer learning modes and an ensemble deep learning algorithm based on stacking were used. The experiments carried out on two datasets of medium and high challenges show very interesting results with F-score reaching 99% for the classification of image modalities and 98% for the classification of anatomical organs.
Offline Arabic handwritten character recognition: from conventional machine learning system to deep learning approaches
(2022) Faouci, Soumia; Gaceb, Djamel; Haddad, Mohammed
Researchers have made great strides in the area of Arabic handwritten character recognition in the last decades especially with the fast development of deep learning algorithms. The characteristics of Arabic manuscript text pose several problems for a recognition system. This paper presents a conventional machine learning system based on the extraction of a set of preselected features and an SVM classifier. In the second part, a simplified convolutional neural network (CNN) model is proposed, which is compared to six other CNN models based on the pre-trained architectures. The suggested methods were tested using three databases: two versions of the OIHACDB dataset and the AIA9K dataset. The experimental results show that the proposed CNN model obtained promising results, as it is able to recognise 94.7%, 98.3%, and 95.6% of the test set of the three databases OIHACDB-28, OIHACDB-40, and AIA9K, respectively.
Progressive Deep Transfer Learning for Accurate Glaucoma Detection in Medical Imaging
(Institute of Electrical and Electronics Engineers, 2024) Yakoub, Assia; Gaceb, Djamel; Touazi, Fayçal; Bourahla, Nourelhouda
Glaucoma leads to permanent vision disability by damaging the optic nerve, which transmits visual images to the brain. The fact that glaucoma doesn't exhibit any symptoms as it progresses and can't be halted in later stages makes early diagnosis critical. Although various deep learning models have been applied to detect glaucoma from digital fundus images, the scarcity of labeled data has limited their generalization performance, along with their high computational complexity and specialized hardware requirements. In this study, a progressive transfer learning with preprocessing techniques is proposed for the early detection of glaucoma in fundus images. The performance of this approach is compared against transfer learning and convolutional neural networks using three benchmark datasets: Cataract, Glaucoma and Origa. The experimental results demonstrate that reusing pre-trained models from ImageNet and applying them to a database containing the same disease leads to improved performance, compared to using databases with different diseases in progressive transfer learning. Additionally, applying preprocessing techniques to the databases further enhances the results.
A review of recent progress in deep learning-based methods for MRI brain tumor segmentation
(Institute of Electrical and Electronics Engineers Inc, 2020) Chihati, S.; Gaceb, Djamel
Brain tumor segmentation is a challenging task that involves delimiting cancerous tissues with heterogeneous and diffuse forms in brain medical images. This process is undoubtedly an important step in computer-aided diagnosis systems, in which tumor regions must be isolated for visualization and subsequent analysis. Recently, great progress has been made in brain tumor segmentation with the emergence of deep learning-based methods, which automatically learn hierarchical, and discriminative features from raw data. These methods outperformed the classical machine learning approaches where handcrafted features are used to describe the differences between pathological and healthy tissues. In this paper, we present a comprehensive overview of recent progress in deep learning-based methods for brain tumor segmentation from magnetic resonance images. Moreover, we discuss the most common challenges and suggest possible solutions
Review on deep learning optimization using knowledge and dataset distillation in medical imaging diagnostics
(2025) Laribi, Nor-Elhouda; Gaceb, Djamel; Rezoug, Abdellah; Touazi, Faycal
The integration of deep learning-based artificial intelligence solutions in hospital environments introduces significant challenges, including data privacy restrictions, limited computational resources, and constraints related to the quality and simplicity of the models used. In this review, we highlight the recent advancements in knowledge distillation and dataset distillation as emerging solutions to these challenges in the field of medical imaging. These techniques offer practical benefits in clinical settings by enabling faster training, reduced model size, improved inference speed, and enhanced accuracy, while supporting privacy-preserving learning across decentralized systems and edge devices. Knowledge distillation transfers knowledge from a complex to a simple model, enabling efficient deployment without high loss in diagnostic performance. Dataset distillation, by contrast, focuses on synthesizing datasets that match the pretrained model on real data, reducing data storage requirements. Together, these methods improve learning efficiency, model accuracy, and resource optimization in hospital workflows. However, their integration into medical environments also presents limitations. Challenges such as pipeline complexity, scalability issues, and performance inconsistency across architectures or high-resolution tasks still persist. Overall, this review provides a comprehensive overview of potential and limitations of these two types of distillations in healthcare, offering insights into how these methods can support more scalable, accurate, and privacy-aware AI solutions for medical imaging.
Super-resolution of document images using transfer deep learning of an ESRGAN model
(IEEE, 2022) Kezzoula, Zakia; Gaceb, Djamel; Gritli, Nadjat
This paper presents a novel super-resolution approach of document images. It is based on transfer deep learning of an ESRGAN model. This model, which showed good robustness on natural images, has been adapted to document images by using better levels of fine-tuning and a post-processing to enhance contrast. The experiments were carried out on our document image dataset that we built from document images presenting different challenges. Documents of different categories with different complexity levels and degradation kinds. The results obtained are better compared to ten existing approaches, which we have developed and tested on the same dataset with the same evaluation protocol
Two-Stage approach for semantic image segmentation of breast cancer: deep learning and mass detection in mammographic images
(CEUR Workshop Proceedings, 2023) Touazi, Faycal; Gaceb, Djamel; Chirane, Marouane; Herzallah, Selma
Breast cancer is a significant global health problem that predominantly affects women and requires effective screening methods. Mammography, the primary screening approach, presents challenges such as radiologist workload and associated costs. Recent advances in deep learning hold promise for improving breast cancer diagnosis. This paper focuses on early breast cancer detection using deep learning to assist radiologists, reduce their workload and costs. We employed the CBIS-DDSM dataset and various CNN models, including YOLO versions V5, V7, and V8 for mass detection, and transformer-based (nested) models inspired by ViT for mass segmentation. Our diverse approach aims to address the complexity of breast cancer detection and segmentation from medical images. Our results show promise, with a 59% mAP50 for cancer mass detection and an impressive 90.15% Dice coefficient for semantic segmentation. These findings highlight the potential of deep learning to enhance breast cancer diagnosis, paving the way for more efficient and accurate early detection methods.
Word-Spotting approach using transfer deep learning of a CNN network
(IEEE, 2020) Benabdelaziz, Ryma; Gaceb, Djamel; Haddad, Mohammed
Convolutional Neural Networks (CNNs) are deep learning models that are trained to automatically extract the most discriminating features directly from an input image to be used for visual classification tasks. Recently, CNNs attracted a lot of interest thanks to their effectiveness in many computer vision applications (medical imaging, video surveillance, biometrics, pattern recognition, OCR, etc.). Transfer learning is an optimization method that uses a pretrained network to speed up the training of another related task or application. This helps speed up and improve the training process on a new dataset. In this paper, we propose a new approach of handwritten word retrieval based on deep learning and transfer learning. We compared the performance between two types of extracted features based on transfer learning: from a pre-trained model and a fine-tuned network. Experiments are performed using six different CNN architectures and three similarity measures on the presegmented Bentham dataset of the ICDAR competition. The obtained results demonstrate the effectiveness of our proposed approach compared to existing methods, evaluated in this competition