Using Fine-tuned Deep Neural Networks for Automated Diagnosis of COVID-19 with Limited Posteroanteriorchest X-ray Images
Introduction:
The paper, published in the Journal of Springer Applied Intelligence in 2020, presents a deep learning-based computer-aided diagnosis system for classifying COVID-19 infected samples using chest X-ray imaging. Due to limited availability of COVID-19 samples, random oversampling and weighted class loss function approaches were proposed to achieve unbiased fine-tuned learning. The dataset used in the study combines COVID-19 samples with CXR images from the RSNA and NLM(MC) datasets. State-of-the-art deep learning approaches such as ResNet, Inception-v3, Inception ResNet-v2, DenseNet169, and NASNetLarge were considered for the classification task. The overall training framework, including the use of transfer learning, is depicted in Fig. 1. The results obtained using the NASNetLarge model and the LIME visualization technique are shown in Fig. 2, providing insights into the classification process. The study concludes by highlighting the different performance of the models based on the classification scenario and suggesting future research directions. More information can be found in the published paper.
Full Article: Using Fine-tuned Deep Neural Networks for Automated Diagnosis of COVID-19 with Limited Posteroanteriorchest X-ray Images
Deep-Learning System for COVID-19 Classification Using Chest X-ray Imaging
Introduction
The outbreak of the novel coronavirus 2019 (COVID-19) has resulted in a global pandemic. Given the highly contagious nature of the virus, there is a need for effective methods of diagnosis. To address this, researchers have proposed a deep learning-based computer-aided diagnosis system that utilizes chest X-ray imaging to classify COVID-19 infected samples.
Proposed Method
To overcome the limited availability of COVID-19 samples, the researchers implemented random oversampling and a weighted class loss function approach for unbiased fine-tuned learning (transfer learning). Additionally, the dataset was augmented with randomly selected CXR images from another dataset, namely RSNA and NLM(MC). The proposed dataset can be accessed on the researcher’s GitHub page.
Deep Learning Approaches
Various state-of-the-art deep learning approaches, including ResNet, Inception-v3, Inception ResNet-v2, DenseNet169, and NASNetLarge, were considered for COVID-19 classification. The problem was divided into binary classification (normal vs. COVID-19 cases) and multi-class classification (COVID-19, pneumonia, and normal cases) of posteroanterior CXR images.
Overview of the Proposed Framework
The proposed framework involves transfer learning on pre-trained models, followed by fine-tuning the head layers to classify COVID-19 samples. The network’s head layers, along with fully connected layers, undergo trainable parameter adjustment. The number of neurons in the output layer depends on whether it’s a binary or multi-class classification. The classified results are then visualized using local interpretable model-agnostic explanations (LIME).
Classification Results and Visualization
The COVID-19 classification results using the NASNetLarge model and LIME visualization are shown in Figure 2. The LIME explanation provides insights into the prediction probabilities of sample being normal, COVID-19, or other pneumonia.
Findings and Conclusion
The deep learning models displayed varying performance depending on the scenario (binary/multi-class classification and oversampling/weighted loss). NASNetLarge exhibited superior performance, particularly in binary classification of COVID-19 samples. The use of local interpretable model-agnostic explanations helped understand the model’s prediction basis. Future research may explore additional deep learning models and preprocessing techniques to further improve results.
For more detailed information on this study, please refer to the original research paper [link here].
Summary: Using Fine-tuned Deep Neural Networks for Automated Diagnosis of COVID-19 with Limited Posteroanteriorchest X-ray Images
The paper titled “Deep Learning-Based Computer-Aided Diagnosis System for Classifying COVID-19 Infected Samples Using Chest X-Ray Imaging” was published in the journal of Springer Applied Intelligence in 2020. The study proposes a deep learning approach to classify COVID-19 infected samples based on chest X-ray imaging. To address the limited availability of COVID-19 samples, the study uses random oversampling and a weighted class loss function. Various state-of-the-art deep learning approaches are considered, and a new dataset is proposed. The framework is fine-tuned using transfer learning, and the results are visualized using local interpretable model-agnostic explanations. The findings show that NASNetLarge performs well in binary classification of COVID-19 samples. For more information, please refer to the published paper.
Frequently Asked Questions:
1. Question: What is deep learning and how does it differ from traditional machine learning?
Answer: Deep learning is a subfield of machine learning that focuses on artificial neural networks and their ability to learn from large amounts of data. Unlike traditional machine learning, deep learning involves training layers of interconnected neural networks to extract and process increasingly complex patterns and features. This approach enables deep learning models to automatically learn hierarchical representations, making them highly effective in solving complex tasks like image recognition and natural language processing.
2. Question: What are the key advantages of using deep learning in various applications?
Answer: Deep learning offers several advantages that make it a preferred choice in various applications. Firstly, it excels at automatically learning intricate patterns and features from raw data, eliminating the need for manual feature engineering. Secondly, deep learning models can handle massive amounts of data, enabling more accurate predictions and insights. Additionally, deep learning networks are highly flexible and adaptable to different domains, making them suitable for a wide range of tasks. Finally, deep learning models continually improve their performance with more data, leading to continuous enhancements and better results over time.
3. Question: What are the main challenges associated with implementing deep learning?
Answer: While deep learning has revolutionized many fields, it is not without challenges. One major challenge is the need for extensive computational resources, including powerful GPUs, to train and run deep neural networks effectively. Deep learning models can be computationally intensive and require significant memory and processing capabilities. Another challenge is the availability and quality of labeled training data, as deep learning models often require large amounts of annotated data for robust training. Additionally, understanding and explaining the decisions made by deep learning models, also known as interpretability, is an ongoing challenge in the field.
4. Question: Are there any limitations or potential risks associated with deep learning?
Answer: Deep learning, like any other technology, has limitations and potential risks. One limitation is the overreliance on training data; if the training data is biased or incomplete, deep learning models may produce biased or inaccurate results. Additionally, deep learning models can be susceptible to adversarial attacks, where subtle alterations to input data can cause the model to make erroneous predictions. Another concern is the lack of interpretability, as deep learning models often operate as black boxes, making it difficult to understand why they made a particular decision. Finally, deep learning models require substantial computational power, which may limit their deployment on resource-constrained devices.
5. Question: How can deep learning be applied in real-world scenarios?
Answer: Deep learning has found applications in various real-world scenarios. In healthcare, deep learning models have been used for medical imaging analysis, disease diagnosis, and drug discovery. In autonomous vehicles, deep learning enables object recognition, lane detection, and decision-making capabilities. Deep learning also plays a significant role in natural language processing applications, such as speech recognition, machine translation, and chatbots. Moreover, industries like finance, e-commerce, and cybersecurity utilize deep learning for fraud detection, recommendation systems, and anomaly detection. The potential applications of deep learning are vast and continually expanding as the technology advances.
