Nancy M. Salem
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Nancy M. Salem has a total of 33 co-authors. Here are the top co-authors by number of shared publications:
- Walid Al-Atabany
Nile University
6 shared publications - Ahmed Farag Seddik
Biomedical Engineering Department, Helwan University, Cairo, Egypt
4 shared publications - Asoke K. Nandi
Department of Electrical Engineering and Electronics, University of Liverpool, Brownlow Hill, Liverpool L69 3GJ, UK
3 shared publications - Hassan M. Ahmed
Helwan University
3 shared publications - Sameh Abd El-Rahman Salem
Helwan University
3 shared publications A.K. Nandi
2 shared publications
Latest Publications
- Open Access
- 4 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 62121-62136
A Deep Learning-Based Approach for Cell Segmentation in Phase-Contrast Images
Cell segmentation in Phase-Contrast (PhC) microscopy is crucial for non-invasive analysis of live-cell imaging data. However, traditional segmentation methods often struggle with challenges such as low contrast, overlapping cells, and artifacts that are unique to PhC imaging. In this study, we proposed a deep learning-based model utilizing the Mask Regional Convolutional Neural Network (Mask R-CNN) architecture for segmentation of cells in PhC images. To optimize the performance, three optimizers have been evaluated including Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (Adam), and Ranger. The models have been trained and tested on two PhC datasets, PhC-C2DH-U373 and PhC-C2DL-PSC, derived from the ISBI cell tracking challenge. Models were tested with ResNet50, ResNet101, and ViT-B/16 backbones, and ablation studies assessed the effects of backbone architectures and layer fine-tuning strategies. Results demonstrated that CNN-based backbones with ResNet101 achieved higher performance compared to the ResNet50 and ViT-B/16 backbones. The Ranger optimizer outperformed Adam and SGD with an Intersection over Union (IoU) of 94.3% and an F1 score of 97.1% on the PhC-C2DH-U373 dataset, as well as an IoU of 85.3% and F1 score of 92% on the PhC-C2DL-PSC dataset, both with data augmentation. Statistical validation using the Wilcoxon signed-rank test confirmed the significance of these results (P-Value < 0.01). Furthermore, an IoU improvement of over 10% is observed for densely packed cells as in the second dataset compared to traditional and benchmark approaches. Additionally, we introduced a generalized model by combining the two datasets to increase model reliability and reduce computational complexity. The findings highlight the potential of Ranger and the generalized training model to enhance cell segmentation across different microscopy datasets.
- 4 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 13, 62121-62136
Open AccessA Deep Learning-Based Approach for Cell Segmentation in Phase-Contrast Images
Cell segmentation in Phase-Contrast (PhC) microscopy is crucial for non-invasive analysis of live-cell imaging data. However, traditional segmentation methods often struggle with challenges such as low contrast, overlapping cells, and artifacts that are unique to PhC imaging. In this study, we proposed a deep learning-based model utilizing the Mask Regional Convolutional Neural Network (Mask R-CNN) architecture for segmentation of cells in PhC images. To optimize the performance, three optimizers have been evaluated including Stochastic Gradient Descent (SGD), Adaptive Moment Estimation (Adam), and Ranger. The models have been trained and tested on two PhC datasets, PhC-C2DH-U373 and PhC-C2DL-PSC, derived from the ISBI cell tracking challenge. Models were tested with ResNet50, ResNet101, and ViT-B/16 backbones, and ablation studies assessed the effects of backbone architectures and layer fine-tuning strategies. Results demonstrated that CNN-based backbones with ResNet101 achieved higher performance compared to the ResNet50 and ViT-B/16 backbones. The Ranger optimizer outperformed Adam and SGD with an Intersection over Union (IoU) of 94.3% and an F1 score of 97.1% on the PhC-C2DH-U373 dataset, as well as an IoU of 85.3% and F1 score of 92% on the PhC-C2DL-PSC dataset, both with data augmentation. Statistical validation using the Wilcoxon signed-rank test confirmed the significance of these results (P-Value < 0.01). Furthermore, an IoU improvement of over 10% is observed for densely packed cells as in the second dataset compared to traditional and benchmark approaches. Additionally, we introduced a generalized model by combining the two datasets to increase model reliability and reduce computational complexity. The findings highlight the potential of Ranger and the generalized training model to enhance cell segmentation across different microscopy datasets.
- Open Access
Evaluating the Potential of Wearable Technology in Early Stress Detection: A Multimodal Approach
Stress can adversely impact health, leading to issues like high blood pressure, heart diseases, and a compromised immune system. Consequently, using wearable devices to monitor stress is essential for prompt intervention and effective management. This study investigates the efficacy of wearable devices in the early detection of psychological stress, employing both binary and five-class classification models. Significant correlations were observed between stress levels and physiological signals, including Electrocardiogram (ECG), Electrodermal Activity (EDA), and Respiration (RESP), establishing these modalities as reliable biomarkers for stress detection. Utilizing the publicly available Wearable Stress and Affect Detection (WESAD) dataset, we employed two ensemble methods, Majority Voting (MV) and Weighted Averaging (WA), to integrate these signals, achieving maximum accuracies of 99.96% for binary classification and 99.59% for five-class classification. This integration significantly enhances the accuracy and robustness of the stress detection system. Furthermore, ten different classifiers were evaluated, and hyperparameter optimization and K-fold cross-validation ranging from 3-fold to 10-fold were applied. Both time-domain and frequency-domain features were examined separately. A review of commercially available wearable devices supporting these modalities was also conducted, resulting in recommendations for optimal configurations for practical applications. Our findings highlight the potential of multimodal wearable devices in advancing the early detection and continuous monitoring of psychological stress, with significant implications for future research and the development of improved stress detection systems.
Open AccessEvaluating the Potential of Wearable Technology in Early Stress Detection: A Multimodal Approach
Stress can adversely impact health, leading to issues like high blood pressure, heart diseases, and a compromised immune system. Consequently, using wearable devices to monitor stress is essential for prompt intervention and effective management. This study investigates the efficacy of wearable devices in the early detection of psychological stress, employing both binary and five-class classification models. Significant correlations were observed between stress levels and physiological signals, including Electrocardiogram (ECG), Electrodermal Activity (EDA), and Respiration (RESP), establishing these modalities as reliable biomarkers for stress detection. Utilizing the publicly available Wearable Stress and Affect Detection (WESAD) dataset, we employed two ensemble methods, Majority Voting (MV) and Weighted Averaging (WA), to integrate these signals, achieving maximum accuracies of 99.96% for binary classification and 99.59% for five-class classification. This integration significantly enhances the accuracy and robustness of the stress detection system. Furthermore, ten different classifiers were evaluated, and hyperparameter optimization and K-fold cross-validation ranging from 3-fold to 10-fold were applied. Both time-domain and frequency-domain features were examined separately. A review of commercially available wearable devices supporting these modalities was also conducted, resulting in recommendations for optimal configurations for practical applications. Our findings highlight the potential of multimodal wearable devices in advancing the early detection and continuous monitoring of psychological stress, with significant implications for future research and the development of improved stress detection systems.
- 22 October 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 276-279
Classification of Autism Spectrum Disorder using Convolutional Neural Networks from Neuroimaging Data
Current Autism Spectrum Disorder (ASD) diagnosis methods exhibit some limitations as they are based on clinical interviews and observations of behaviors, characteristics, and abilities. Moreover, considering the current challenges in identifying the causes and mechanisms associated with ASD, there is an essential need for automated techniques capable of providing an accurate classification between ASD and typically developed (TD). In this paper, we present a convolutional neural network model that can differentiate ASD from TD. This proposed system is trained and validated on the well-known Autism Brain-Imaging Data Exchange (ABIDE) dataset for the resting-state Functional Magnetic Resonance Imaging (fMRI). Results showed the robustness ofthe proposed system that achieves maximum accuracy of 99%.
- 22 October 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 276-279
Classification of Autism Spectrum Disorder using Convolutional Neural Networks from Neuroimaging Data
Current Autism Spectrum Disorder (ASD) diagnosis methods exhibit some limitations as they are based on clinical interviews and observations of behaviors, characteristics, and abilities. Moreover, considering the current challenges in identifying the causes and mechanisms associated with ASD, there is an essential need for automated techniques capable of providing an accurate classification between ASD and typically developed (TD). In this paper, we present a convolutional neural network model that can differentiate ASD from TD. This proposed system is trained and validated on the well-known Autism Brain-Imaging Data Exchange (ABIDE) dataset for the resting-state Functional Magnetic Resonance Imaging (fMRI). Results showed the robustness ofthe proposed system that achieves maximum accuracy of 99%.
- 1 October 2021
- Published by Wolters Kluwer Health in Journal of Clinical Engineering
- Vol. 46 (4) , 177-188
Impact of Post–Refractive Surgeries on Corneal Biomechanics—A Review
A stable shape for the human cornea prevents complications of post–refractive surgeries. Variation in corneal lamellae and collagen fibers is induced by recent laser refractive surgical procedures utilizing laser ablation and disruption techniques. Noninvasive prediction of corneal biomechanics and structural strength is not quite accurate pre– and post–refractive surgery despite the great advance in assessment methods. This is due to the alteration happening to the corneal structure post–refractive surgery. Many different methods have been developed to measure corneal and subcorneal biomechanics. In this article, an overview of the impact of post–refractive surgeries on corneal biomechanics is presented. The corneal anatomy and physiology are reviewed. Refractive disorders, their corresponding surgeries, and their relation to corneal biomechanics are discussed as well. Ultrasonic estimation of corneal biomechanics and their limitations are stated in this overview.
- 1 October 2021
- Published by Wolters Kluwer Health in Journal of Clinical Engineering
- Vol. 46 (4) , 177-188
Impact of Post–Refractive Surgeries on Corneal Biomechanics—A Review
A stable shape for the human cornea prevents complications of post–refractive surgeries. Variation in corneal lamellae and collagen fibers is induced by recent laser refractive surgical procedures utilizing laser ablation and disruption techniques. Noninvasive prediction of corneal biomechanics and structural strength is not quite accurate pre– and post–refractive surgery despite the great advance in assessment methods. This is due to the alteration happening to the corneal structure post–refractive surgery. Many different methods have been developed to measure corneal and subcorneal biomechanics. In this article, an overview of the impact of post–refractive surgeries on corneal biomechanics is presented. The corneal anatomy and physiology are reviewed. Refractive disorders, their corresponding surgeries, and their relation to corneal biomechanics are discussed as well. Ultrasonic estimation of corneal biomechanics and their limitations are stated in this overview.
- Open Access
- 26 July 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 106014-106028
Corneal Biomechanics Assessment Using High Frequency Ultrasound B-Mode Imaging
Assessment of corneal biomechanics for pre- and post-refractive surgery is of great clinical importance. Corneal biomechanics affect vision quality of human eye. Many factors affect corneal biomechanics such as, age, corneal diseases and corneal refractive surgery. There is a need for non-invasive in-vivo measurement of corneal biomechanics due to corneal shape preserving as opposed to ex-vivo measurements that destructs corneal tissue. In this study, a new approach for assessing corneal biomechanics in-vivo non-invasively using ultrasound estimation method with 100 KHz frame rate is proposed. Three models in conjunction with each other are used to study the biomechanics behavior of corneal tissue pre- and post-refractive surgery. These three models are cornea FEM, cornea scatterrer model and ultrasound transducer model respectively. Nine different elastic moduli corneal models are constructed to achieve this goal; 140KPa, 300KPa, 600KPa and 800KPa as post-refractive surgery models and 1MPa, 1.5MPa, 2MPa, 2.5MPa and 3MPa as pre-refractive surgery models respectively. Time-To-Peak (TTP) deformation, deformation amplitude (DA), deformation amplitude ratio at 2 mm (DA ratio 2 mm) and shear wave speed (SWS) are estimated for each of the nine involved corneal models in this study. Results show that, TTP is decreasing (6.9 msec. at 140KPa to 5.3 msec. at 3MPa) while increasing the elastic modulus of corneal tissue. Also, DA is decreasing (2 mm at 140KPa to 0.5 mm at 3MPa), while increasing the elastic modulus as well. However, DA ratio shows decrease while increasing of elastic modulus to reflect the difficulty of corneal tissue to deform uniformly at higher elastic moduli. Estimated SWS shows an average accuracy of 98% of the theoretical SWS.
- 26 July 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 106014-106028
Open AccessCorneal Biomechanics Assessment Using High Frequency Ultrasound B-Mode Imaging
Assessment of corneal biomechanics for pre- and post-refractive surgery is of great clinical importance. Corneal biomechanics affect vision quality of human eye. Many factors affect corneal biomechanics such as, age, corneal diseases and corneal refractive surgery. There is a need for non-invasive in-vivo measurement of corneal biomechanics due to corneal shape preserving as opposed to ex-vivo measurements that destructs corneal tissue. In this study, a new approach for assessing corneal biomechanics in-vivo non-invasively using ultrasound estimation method with 100 KHz frame rate is proposed. Three models in conjunction with each other are used to study the biomechanics behavior of corneal tissue pre- and post-refractive surgery. These three models are cornea FEM, cornea scatterrer model and ultrasound transducer model respectively. Nine different elastic moduli corneal models are constructed to achieve this goal; 140KPa, 300KPa, 600KPa and 800KPa as post-refractive surgery models and 1MPa, 1.5MPa, 2MPa, 2.5MPa and 3MPa as pre-refractive surgery models respectively. Time-To-Peak (TTP) deformation, deformation amplitude (DA), deformation amplitude ratio at 2 mm (DA ratio 2 mm) and shear wave speed (SWS) are estimated for each of the nine involved corneal models in this study. Results show that, TTP is decreasing (6.9 msec. at 140KPa to 5.3 msec. at 3MPa) while increasing the elastic modulus of corneal tissue. Also, DA is decreasing (2 mm at 140KPa to 0.5 mm at 3MPa), while increasing the elastic modulus as well. However, DA ratio shows decrease while increasing of elastic modulus to reflect the difficulty of corneal tissue to deform uniformly at higher elastic moduli. Estimated SWS shows an average accuracy of 98% of the theoretical SWS.
- 1 October 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 1, 214-217
Binary Classification of Visual Scenes Using Convolutional Neural Network
Scene classification is a dominant track in computer vision tasks as it can help in many missions such as navigation, preprocessing, big data organization, albuming systems, and recognition applications for blinds. Recently, Convolutional Neural Network (CNN) shows noteworthy performance in enhancing the results of most image processing research points. In this paper, we use CNN for indoor-outdoor classification problem with the aid of a large database. Inception-v3 model (without its top layers) is used to extract scene features and extra three layers are attached to adopt the classification task. Our approach reaches an overall classification accuracy of 98.4% which shows the robustness of CNN over the old techniques.
- 1 October 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- Vol. 1, 214-217
Binary Classification of Visual Scenes Using Convolutional Neural Network
Scene classification is a dominant track in computer vision tasks as it can help in many missions such as navigation, preprocessing, big data organization, albuming systems, and recognition applications for blinds. Recently, Convolutional Neural Network (CNN) shows noteworthy performance in enhancing the results of most image processing research points. In this paper, we use CNN for indoor-outdoor classification problem with the aid of a large database. Inception-v3 model (without its top layers) is used to extract scene features and extra three layers are attached to adopt the classification task. Our approach reaches an overall classification accuracy of 98.4% which shows the robustness of CNN over the old techniques.
- 1 July 2019
- Published by SAGE Publications in Bio-Medical Materials and Engineering
- Vol. 30 (4) , 449-462
Automated brain tumor segmentation from multi-slices FLAIR MRI images
Brain tumors are considered to be a leading cause of cancer death among young people. Early diagnosis is thus essential for treatment. The brain segmentation process is still challenging due to complexity and variation of the tumor structure, intensity similarity between tumor tissues and normal brain tissues. In this paper, a fully automated and reliable brain tumor segmentation system is proposed. This system is able to detect range of slices from a volume that is likely to contain tumor in MRI images. An iterated k-means algorithm is used for the segmentation process in conjunction with a cluster validity index to select the optimal number of clusters. The proposed approach is evaluated using simulated and real MRI of human brain from multimodal brain tumor image segmentation benchmark (BRATS) organized by MICCAI 2012 challenge. Our results achieved average for Dice overlap and Jaccard index for complete tumor region of 91.96% and 98.31% respectively when testing a set of 77 volumes. This shows the robustness of the new technique for clinical routine use.
- 1 July 2019
- Published by SAGE Publications in Bio-Medical Materials and Engineering
- Vol. 30 (4) , 449-462
Automated brain tumor segmentation from multi-slices FLAIR MRI images
Brain tumors are considered to be a leading cause of cancer death among young people. Early diagnosis is thus essential for treatment. The brain segmentation process is still challenging due to complexity and variation of the tumor structure, intensity similarity between tumor tissues and normal brain tissues. In this paper, a fully automated and reliable brain tumor segmentation system is proposed. This system is able to detect range of slices from a volume that is likely to contain tumor in MRI images. An iterated k-means algorithm is used for the segmentation process in conjunction with a cluster validity index to select the optimal number of clusters. The proposed approach is evaluated using simulated and real MRI of human brain from multimodal brain tumor image segmentation benchmark (BRATS) organized by MICCAI 2012 challenge. Our results achieved average for Dice overlap and Jaccard index for complete tumor region of 91.96% and 98.31% respectively when testing a set of 77 volumes. This shows the robustness of the new technique for clinical routine use.
- Top Cited PapersOpen Access
- 27 May 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 7, 69215-69225
Multi-Classification of Brain Tumor Images Using Deep Neural Network
Brain tumor classification is a crucial task to evaluate the tumors and make a treatment decision according to their classes. There are many imaging techniques used to detect brain tumors. However, MRI is commonly used due to its superior image quality and the fact of relying on no ionizing radiation. Deep learning (DL) is a subfield of machine learning and recently showed a remarkable performance especially in classification and segmentation problems. In this paper, a DL model based on a convolutional neural network is proposed to classify different brain tumor types using two publicly available datasets. The former one classifies tumors into (meningioma, glioma, and pituitary tumor). The other one differentiates between three glioma grades (Grade II, Grade III, and Grade IV). Datasets include 233 and 73 patients with a total of 3064 and 516 images on T1-weighted contrast-enhanced images for the first and second datasets respectively. The proposed network structure achieves a significant performance with a best overall accuracy of 96.13% and 98.7% respectively for the two studies. Results indicate the ability of the model for brain tumor multi-classification purposes.
- 27 May 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 7, 69215-69225
Top Cited PapersOpen AccessMulti-Classification of Brain Tumor Images Using Deep Neural Network
Brain tumor classification is a crucial task to evaluate the tumors and make a treatment decision according to their classes. There are many imaging techniques used to detect brain tumors. However, MRI is commonly used due to its superior image quality and the fact of relying on no ionizing radiation. Deep learning (DL) is a subfield of machine learning and recently showed a remarkable performance especially in classification and segmentation problems. In this paper, a DL model based on a convolutional neural network is proposed to classify different brain tumor types using two publicly available datasets. The former one classifies tumors into (meningioma, glioma, and pituitary tumor). The other one differentiates between three glioma grades (Grade II, Grade III, and Grade IV). Datasets include 233 and 73 patients with a total of 3064 and 516 images on T1-weighted contrast-enhanced images for the first and second datasets respectively. The proposed network structure achieves a significant performance with a best overall accuracy of 96.13% and 98.7% respectively for the two studies. Results indicate the ability of the model for brain tumor multi-classification purposes.
- 1 February 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 35-40
Investigating Relationship Between Lateral Time-To-Peak (TTP) Displacement Curves and Stiffness of Viscoelastic Agar-Gelatin Phantoms
Shear wave speed is a metric for the tissue elastic properties and its pathologic changes. This is due to the implicit relationship between the speed of the propagating shear wave within tissue and both its elastic properties and pathologic changes. Wave propagation is a result of a force applied transiently to the tissue model. In this paper, the relationship between lateral TTP displacement curves and stiffness of different Young's moduli phantoms is investigated. Five different elastic moduli Finite Element Models (FEMs) are used; 5.2KPa, 9.8KPa, 23.9KPa, 44.2KPa and 67.3KPa. Agar-gelatin model is used as a general model for mimicking soft tissue. Displacements are tracked with respect to the simulation time to create the displacement magnitude profile for the central node for each phantom. Results have shown that the second highest peak happens after 0.05 seconds in case of low elastic moduli FEMs. This time stamp can be used as a yardstick to differentiate between low and high elastic moduli without considering estimating the shear wave speed for many lateral nodes.
- 1 February 2019
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 35-40
Investigating Relationship Between Lateral Time-To-Peak (TTP) Displacement Curves and Stiffness of Viscoelastic Agar-Gelatin Phantoms
Shear wave speed is a metric for the tissue elastic properties and its pathologic changes. This is due to the implicit relationship between the speed of the propagating shear wave within tissue and both its elastic properties and pathologic changes. Wave propagation is a result of a force applied transiently to the tissue model. In this paper, the relationship between lateral TTP displacement curves and stiffness of different Young's moduli phantoms is investigated. Five different elastic moduli Finite Element Models (FEMs) are used; 5.2KPa, 9.8KPa, 23.9KPa, 44.2KPa and 67.3KPa. Agar-gelatin model is used as a general model for mimicking soft tissue. Displacements are tracked with respect to the simulation time to create the displacement magnitude profile for the central node for each phantom. Results have shown that the second highest peak happens after 0.05 seconds in case of low elastic moduli FEMs. This time stamp can be used as a yardstick to differentiate between low and high elastic moduli without considering estimating the shear wave speed for many lateral nodes.
- 1 March 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 495-502
Automatic classification of masses from digital mammograms
Cancer of the breast remains to be one among the foremost factors behind death between women. The early detection and recognition of breast cancer increases the probabilities of complete and successful recovery; therefore, the mortality rate can be reduced. Currently, screening mammography is the most successful technique for detecting masses or abnormalities which are related to breast cancer in its earliest stages. In this paper, an automatic method for mass classification from mammograms is introduced. The proposed algorithm includes three main steps. First, three different types of features are separated from the mass. Then the most relevant features are selected using the T-test algorithm. Finally, the classification step is done to distinguish among benign and malignant masses with the aid of the usage of the three classifiers; artificial neural network, support vector machines, k-nearest neighbor. The system is rated using 250 mammograms from the Digital Database for Screening Mammography (DDSM). The artificial neural network occupies the best results of 98.9 % for accuracy, 100 % for sensitivity, and 97.8 % for specificity.
- 1 March 2018
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 495-502
Automatic classification of masses from digital mammograms
Cancer of the breast remains to be one among the foremost factors behind death between women. The early detection and recognition of breast cancer increases the probabilities of complete and successful recovery; therefore, the mortality rate can be reduced. Currently, screening mammography is the most successful technique for detecting masses or abnormalities which are related to breast cancer in its earliest stages. In this paper, an automatic method for mass classification from mammograms is introduced. The proposed algorithm includes three main steps. First, three different types of features are separated from the mass. Then the most relevant features are selected using the T-test algorithm. Finally, the classification step is done to distinguish among benign and malignant masses with the aid of the usage of the three classifiers; artificial neural network, support vector machines, k-nearest neighbor. The system is rated using 250 mammograms from the Digital Database for Screening Mammography (DDSM). The artificial neural network occupies the best results of 98.9 % for accuracy, 100 % for sensitivity, and 97.8 % for specificity.