Noha Medhat Ghatwary , Ph.D.

Assistant Professor



  • Automatic Esophageal Abnormality Detection and Classification

    Esophageal cancer is counted as one of the deadliest cancers worldwide ranking the sixth among all types of cancers. Early esophageal cancer typically causes no symp- toms and mainly arises from overlooked/untreated premalignant abnormalities in the esophagus tube. Endoscopy is the main tool used for the detection of abnormalities, and the cell deformation stage is confirmed by taking biopsy samples. The process of detection and classification is considered challenging for several reasons such as different types of abnormalities (including early cancer stages) can be located ran- domly throughout the esophagus tube, abnormal regions can have various sizes and appearances which makes it difficult to capture, and failure in discriminating between the columnar mucosa from the metaplastic epithelium. Although many studies have been conducted, it remains a challenging task and improving the accuracy of auto- matically classifying and detecting different esophageal abnormalities is an ongoing field. This thesis aims to develop novel automated methods for the detection and classification of the abnormal esophageal regions (precancerous and cancerous) from endoscopic images and videos.
    In this thesis, firstly, the abnormality stage of the esophageal cell deformation is clas- sified from confocal laser endomicroscopy (CLE) images. The CLE is an endoscopic tool that provides a digital pathology view of the esophagus cells. The classifica- tion is achieved by enhancing the internal features of the CLE image, using a novel enhancement filter that utilizes fractional integration and differentiation. Different imaging features including, Multi-Scale pyramid rotation LBP (MP-RLBP), gray level co-occurrence matrices (GLCM), fractal analysis, fuzzy LBP and maximally stable extremal regions (MSER), are calculated from the enhanced image to assure a robust classification result. The support vector machine (SVM) and random forest (RF) classifiers are employed to classify each image into its pathology stage.
    Secondly, we propose an automatic detection method to locate abnormality regions from high definition white light (HD-WLE) endoscopic images. We first investigate the performance of different deep learning detection methods on our dataset. Then we propose an approach that combines hand-designed Gabor features with extracted convolutional neural network features that are used by the Faster R-CNN to detect abnormal regions. Moreover, to further improve the detection performance, we pro- pose a novel two-input network named GFD-Faster RCNN. The proposed method generates a Gabor fractal image from the original endoscopic image using Gabor filters. Then features are learned separately from the endoscopic image and the gen- erated Gabor fractal image using the densely connected convolutional network to detect abnormal esophageal regions.
    Thirdly, we present a novel model to detect the abnormal regions from endoscopic videos. We design a 3D Sequential DenseConvLstm network to extract spatiotem- poral features from the input videos that are utilized by a region proposal network and ROI pooling layer to detect abnormality regions in each frame throughout the video. Additionally, we suggest an FS-CRF post-processing method that incorpor- ates the Conditional Random Field (CRF) on a frame-based level to recover missed abnormal regions in neighborhood frames within the same clip.
    The methods are evaluated on four datasets: (1) CLE dataset used for the classific- ation model, (2) Publicly available dataset named Kvasir, (3) MICCAI’15 Endovis challenge dataset, Both datasets (2) and (3) are used for the evaluation of detection model from endoscopic images. Finally, (4) Gastrointestinal Atlas dataset used for the evaluation of the video detection model. The experimental results demonstrate promising results of the different models and have outperformed the state-of-the-art methods.

List of Publications

Address :

Phone : +(20 3) 5622366

Room No: 0

Email: Send Mail

Web page : link