Only 14 pages are availabe for public view
Early diagnosis of ophthalmic tumors and their classification is an extremely necessary topic in the ophthalmology field. Ophthalmic tumors are rare and diverse so their diagnosis can be quite complex. Treatment usually requires special expertise and equipment and in many instances is controversial. Quality-of-life measures have become increasingly important in ophthalmic oncology which increases the survival rate. Assisting and supporting physicians in identifying ocular tumors require an efficient intelligence system with high quality to classify Eye Tumors. For classification, some previous studies have applied machine learning techniques however other studies have applied Artificial intelligence (AI) techniques. Furthermore, merging the skills of human physicians with the powers of deep learning algorithms should decrease diagnostic and treatment mistakes. So, this research conducted the development of intelligent systems using deep learning techniques to classify and identify eye tumors. These systems applied different architectures of convolutional neural networks (CNN) with different datasets.
The main objective of this research is to study the characteristics and features of each eye tumor and then interpret it, using it to develop a new intelligent system to diagnose and identify the eye tumor types from medical images. This research is to study the benefits and drawbacks of the related work and their models to determine the efficient techniques to apply. Also, this research proved that deep learning techniques especially convolutional neural networks (CNN) is a more effective method than traditional techniques. Deep learning algorithms have outperformed machine learning algorithms and have shown to be the best tools for a wide range of medical applications. It has the ability to learn features from a large amount of data and self-correct to enhance accuracy. Moreover, CNN is the most remarkable technique in supervised learning. It is considered a feature extraction procedure that detects essential features in data with a grid pattern, such as images. Because of the rarity and limitation of eye tumors datasets, the existence of intelligent models for their classification is few, and for specific eye tumors, such as orbital tumor classification, no related studies have been proposed before. As a result, it is a motivating research area for developing a novel model that performs well while working with imbalanced datasets.
Therefore, this research implemented to design of three intelligent systems by using deep learning techniques, especially convolutional neural networks to diagnose and classify eye tumors while solving the imbalanced datasets problem. Each system is developed to identify one of the malignancy eye tumors; the first system is used to diagnose retinoblastoma while the second and third systems are to diagnose orbital tumors. The datasets used in each system are different imbalanced datasets, collected from two locations. The tumor images are collected from Ain Shams University Hospital, Ophthalmology department while the normal images are obtained from public websites like Kaggle and medical images websites. The first system consists of four main stages: pre-processing the image dataset, re-sampling, and data augmentation of the image dataset, developing the training model, and finally evaluating the model to classify retinoblastoma.
The first system has applied five preprocessing image techniques to enhance and increase the image dataset. Then, it used oversampling as the second method and perfect re-sampling techniques to balance the input image dataset in the model. It has applied four image augmentation techniques. A convolutional neural network was applied as a feature extraction and classifier for the model. The experiment used different epochs and batch sizes to train the model. The best experimental result of the first system demonstrated 80 epochs and 64 of batch size and evaluation metrics accuracy, precision, recall, and f1-score are 98%, 98.3%, 97%, and 98%, respectively.
The second system consists of four stages: pre-processing of the image datasets, re-sampling and data augmentation of the image datasets, development of the training model, and lastly evaluation of the model to identify orbital tumors. It used two different datasets of magnetic resonance imaging (MRI) which are T1-weighted and T2-weighted images to have the ability to compare the experimental results because no previous studies were proposed. It used eight image pre-processing methods to improve and enlarge the image collection. The model’s input image dataset was then balanced using oversampling as the second way and excellent re-sampling techniques. It used seven image augmentation methods. The model’s features extractor and classifier used a convolutional neural network. Several epochs and batch sizes were utilized in the experiment to train the model. The most significant experimental results of the second system displayed in 80 epochs and 64 batch sizes, as well as the evaluation metrics accuracy, precision, recall, and f1-score, are 95%, 95%, 94%, and 96% for T1-weighted images, respectively and 94%, 91%, 90% and 94% for T2-weighted images, respectively.
The third system consists of four main stages: image dataset pre-processing, image dataset re-sampling and data augmentation, training model creation, and model evaluation to identify orbital tumors. It employed identical magnetic resonance imaging (MRI) images as the second experiment, which are T1-weighted and T2-weighted images. To improve and expand the image collection, it applied eight pre-processing image approaches. After that, the model’s input image dataset was balanced by utilizing oversampling as the second method and good re-sampling techniques. It applied seven image augmentation techniques. As the model’s features extractor and classifier, a convolutional neural network was deployed. To train the model, many epochs and batch sizes were utilized in the experiment. Moreover, the third system’s major objective was to develop a novel convolutional neural network model with the best feature extraction and to reduce the dimensionality of the feature map in order to improve system classification performance. The third system improved the model’s characteristics and ocular tumor recognition by using five convolutional layers as feature extraction in a convolutional neural network. According to accuracy measurements, the most significant experimental results of the third system demonstrated improved efficacy and performance of the proposed system. The results are shown in 80 epochs and 64 batch sizes, and the evaluation metrics accuracy, precision, recall, and f1-score are 98%, 99%, 97%, and 99% for T1-weighted images, respectively, and 97%, 98%, 94.8%, and 97% for T2-weighted images.