Browsing by Author "Haftu, Kibrom"

Now showing 1 - 4 of 4
  • No Thumbnail Available
    Item
    Cardiorespiratory Disease Detection Using Deep Transfer Learning
    (Addis Ababa University, 2021-02-25) Haftu, Kibrom; Assabie, Yaregal (PhD)
    Cardiorespiratory diseases are recognized as serious, worldwide public health concerns that have remained among the leading cause of death globally. These diseases range from common colds to life-threatening bacterial pneumonia that affects anyone, anywhere. Timely, accurate, and effective detection is critical, especially for infectious diseases to prevent the condition from becoming too severe. Deep learning a subfield of machine learning played a central role to detect cardiorespiratory diseases from X-ray images using transfer learning techniques. These approaches allow radiologists (or other health professionals) to extract relevant features that are undetectable to the naked eye or difficult to spot. However, there are still longstanding issues related to the application of deep learning for transfer learning. Difficulties arise when transferring knowledge from one domain into another task that could hurt model performance. We, therefore, proposed a non-parametric Bayesian statistical model to investigate the effectiveness of transfer learning on X-ray images. The proposed model comprises two main components: The deep transfer learning and the detection component. The first component leverages unlabeled X-ray images using unsupervised variational autoencoder to EfficientNet architecture. Its purpose is to extract robust features for detection. The detection component further fine-tunes these features using an end-to-end supervised algorithm rather than current approaches that use a single feature space represented by the last fully connected layer of the convolutional neural network across all conditions. At this stage, we conducted a hyperparameter search for different parameters. Moreover, we developed an offline data augmentation algorithm to address the class imbalance problem that emerged from the dataset. In doing so, we expand the training data from 75,682 to 108,708 X-ray images The proposed system was implemented on Google Cloud Platform using an open-source deep learning library called PyTorch. The National Institutes of Health dataset was used to train, validate, and test the proposed model. The dataset is comprising 112, 120 images from nearly 31,000 patients. We also evaluated the performance of our approaches using Tikur Anbessa Specialized Hospital X-ray images. We have achieved an 88.01% AUC score that outperforms the existing state-of-the-art approaches.
  • No Thumbnail Available
    Item
    Cardiorespiratory Disease Detection Using Deep Transfer Learning
    (Addis Ababa University, 2/25/2021) Haftu, Kibrom; Assabie, Yaregal (PhD)
    Cardiorespiratory diseases are recognized as serious, worldwide public health concerns that have remained among the leading cause of death globally. These diseases range from common colds to life-threatening bacterial pneumonia that affects anyone, anywhere. Timely, accurate, and effective detection is critical, especially for infectious diseases to prevent the condition from becoming too severe. Deep learning a subfield of machine learning played a central role to detect cardiorespiratory diseases from X-ray images using transfer learning techniques. These approaches allow radiologists (or other health professionals) to extract relevant features that are undetectable to the naked eye or difficult to spot. However, there are still longstanding issues related to the application of deep learning for transfer learning. Difficulties arise when transferring knowledge from one domain into another task that could hurt model performance. We, therefore, proposed a non-parametric Bayesian statistical model to investigate the effectiveness of transfer learning on X-ray images. The proposed model comprises two main components: The deep transfer learning and the detection component. The first component leverages unlabeled X-ray images using unsupervised variational autoencoder to EfficientNet architecture. Its purpose is to extract robust features for detection. The detection component further fine-tunes these features using an end-to-end supervised algorithm rather than current approaches that use a single feature space represented by the last fully connected layer of the convolutional neural network across all conditions. At this stage, we conducted a hyperparameter search for different parameters. Moreover, we developed an offline data augmentation algorithm to address the class imbalance problem that emerged from the dataset. In doing so, we expand the training data from 75,682 to 108,708 X-ray images The proposed system was implemented on Google Cloud Platform using an open-source deep learning library called PyTorch. The National Institutes of Health dataset was used to train, validate, and test the proposed model. The dataset is comprising 112, 120 images from nearly 31,000 patients. We also evaluated the performance of our approaches using Tikur Anbessa Specialized Hospital X-ray images. We have achieved an 88.01% AUC score that outperforms the existing state-of-the-art approaches.
  • No Thumbnail Available
    Item
    Tigrigna Question Answering System for Factoid Questions
    (Addis Ababa University, 6/17/2016) Haftu, Kibrom; Assabie, Yaregal (PhD)
    Accessing relevant information is one of the major problems faced by Tigrigna language users for every domain of knowledge when dealing with huge amount of information especially in the Internet. Evidently, users are interested in obtaining a specific and precise answer to a specific question. However, obtaining a relevant and concise answer is a challenge to particular user question. For such situation, Tigrigna Question Answering system is a good solution. The proposed QA system comprises of question analysis, document analysis and answer extraction modules. The main function of question analysis module is taking a Tigrigna Question as input and then generates a query, expands a query and determines its Question Particle and Question Type. A statistical language model approach is used to model the classification of Tigrigna questions to their category or type. The document analysis module performs the process of pre-processing of parallel corpora, which are documents that contain question sentences in one document and answer sentences in another one, and also ranking and extracting answer contents. Answer extraction also performs the detail analysis on the retrieved answer contents based on the question type, question particle and query using the techniques of language modeling called Answer Model. This statistical language model does the extraction process of exact and precise Tigrigna answer in probabilistic manner from sets candidate answers. Generally, this system developed after reviewed literatures and related work, and selected the appropriate tools and data source such as Moses, GIZA++ and IRSTLM as tools and different Webs and Tigrigna newspapers and magazines as data sources. Our data sets are classified for training and testing activities of the system. Based on this, we collected around 1000 data sets for training and 200 data sets for testing. Performance evaluation conducted manually by comparing the system‟s answers with the answers exists in testing document, which is prepared for testing purpose. Finally the evaluation results of Tigrigna factoid QAS is expressed in terms of the average performance of a question type classifier which is 87%, and the average Precision, Recall and F – measure of the answer extraction, precision is 88.5%, recall is 85.9% and F – measure is 87.2%. Keywords: Tigrigna question answering, Tigrigna Factoid questions, Language model based question classification, question analysis, Document Analysis, Answer Extraction.
  • No Thumbnail Available
    Item
    Tigrigna Question Answering System for Factoid Questions
    (Addis Ababa University, 2016-06-17) Haftu, Kibrom; Assabie, Yaregal (PhD)
    Accessing relevant information is one of the major problems faced by Tigrigna language users for every domain of knowledge when dealing with huge amount of information especially in the Internet. Evidently, users are interested in obtaining a specific and precise answer to a specific question. However, obtaining a relevant and concise answer is a challenge to particular user question. For such situation, Tigrigna Question Answering system is a good solution. The proposed QA system comprises of question analysis, document analysis and answer extraction modules. The main function of question analysis module is taking a Tigrigna Question as input and then generates a query, expands a query and determines its Question Particle and Question Type. A statistical language model approach is used to model the classification of Tigrigna questions to their category or type. The document analysis module performs the process of pre-processing of parallel corpora, which are documents that contain question sentences in one document and answer sentences in another one, and also ranking and extracting answer contents. Answer extraction also performs the detail analysis on the retrieved answer contents based on the question type, question particle and query using the techniques of language modeling called Answer Model. This statistical language model does the extraction process of exact and precise Tigrigna answer in probabilistic manner from sets candidate answers. Generally, this system developed after reviewed literatures and related work, and selected the appropriate tools and data source such as Moses, GIZA++ and IRSTLM as tools and different Webs and Tigrigna newspapers and magazines as data sources. Our data sets are classified for training and testing activities of the system. Based on this, we collected around 1000 data sets for training and 200 data sets for testing. Performance evaluation conducted manually by comparing the system‟s answers with the answers exists in testing document, which is prepared for testing purpose. Finally the evaluation results of Tigrigna factoid QAS is expressed in terms of the average performance of a question type classifier which is 87%, and the average Precision, Recall and F – measure of the answer extraction, precision is 88.5%, recall is 85.9% and F – measure is 87.2%. Keywords: Tigrigna question answering, Tigrigna Factoid questions, Language model based question classification, question analysis, Document Analysis, Answer Extraction.