Neural Network Can Diagnose Covid-19 from Chest X-Rays
- New study is 98.4% accurate at detecting Covid-19 from X-rays.
- Researchers trained a convolutional neural network on Kaggle dataset.
- The hope is that the technology can be used to quickly and effectively identify Covid-19 patients.
As the Covid-19 pandemic continues to evolve, there is a pressing need for a faster diagnostic system. Testing kit shortages, virus mutations, and soaring numbers of cases have overwhelmed health care systems worldwide. Even when a good testing policy is in place, lab testing is arduous, expensive, and time consuming. Cheap antigen tests, which can give results in 30 seconds, are widely available but suffer from low sensitivity; The tests correctly identifying just 75% of Covid-19 cases a week after symptoms start .
Shashwat Sanket and colleagues set out to find an easy, fast, and accurate alternative using simple chest X-ray images. The team found that bilateral changes seen in chest X-rays of patients with Covid-19 can be analyzed and classified without a radiologist’s interpretation, using Convolutional Neural Networks (CNNs). The study, published in the September issue of Multimedia tools and Applications, successfully trained a CNN to accurately diagnose Covid-19 from Chest X-Rays, achieving an impressive 98.4% classification accuracy.. The journal article, titled Detection of novel coronavirus from chest X-rays using deep convolutional neural networks, shows some exciting promise in the ongoing efforts to find ways to detect Covid-19 quickly and effectively,
What are Convolutional Neural Networks?
A convolutional neural network (CNN) is a Deep Learning algorithm that resembles the response of neurons in the visual cortex. The algorithm takes an input image and weighs the relative importance of various aspects in the image. The neurons overlap to span the entire field of vision, comprising a completely connected network where neurons in one layer link to neurons in other layers. The multilayered CNN includes an input layer, an output layer, and several hidden layers. A simple process called pooling keeps the most important features while reducing the dimensionality of the feature map.
One major advantage of CNNs is that, compared to other classification algorithms, the required pre-processing is much lower. In addition, CNNs use regularized weights over fewer parameters. This avoids the exploding gradient and vanishing gradient problems of traditional neural networks during backpropagation.
The study began with a Kaggle dataset containing radiography images. As well as chest X-ray images for 219 COVID-19 positive cases, the dataset also contained 1341 normal chest X-rays and 1345 viral pneumonia images. Random selection was used to reduce the normal and viral pneumonia images to a balanced 219 each. The model, which the authors dubbed CovCNNl, was trained with augmented chest X-ray images; The raw images were standardized with each other using transformations like shearing, shifting and rotation. They were also converted to the same size: 224 × 224 × 3 pixels. Following the augmentation, the dataset was split into 525 images for training and 132 images for testing. The following image, from the study authors, demonstrates how the augmented images appear. Image a in the top row shows how Covid-19 appears on an x-ray, in comparison to four normal chest X-rays:
Seven existing pre-trained transfer learning models were used in the study, including ResNet-101 (a 101 layers deep CNN), Xception (71 layers deep), and VGG-16, which is widely used in image classification problems but painfully slow to train . Transfer learning takes lessons learned from previous classification problems and transfers that knowledge to a new task—in this case, correctly identifying COVID-19 patients.
Four variant CovCNN models were tested for effectiveness with several metrics, including: accuracy, F1-score, sensitivity, and specificity. The F1 score is a combination of recall and precision; Sensitivity is the true positive rate—the proportion of correctly predicted positive cases; Specificity is the proportion of correctly identified negative cases. The CovCNN_4 model outperformed all the other models, achieving 98.48% accuracy, 100% sensitivity, and 97.73% specificity. This fine-tuned deep network contained 15 layers, stacked sequentially with increasing filter sizes. This next image shows the layout of the model:
The authors conclude that their covCNN_4 model can be employed to assist medical practitioners and radiologists with faster, more accurate Covid-19 diagnosis, as well as follow up cases. In addition, they recommend that their model’s accuracy can be further improved by “fusion of CNN and pre-trained model features”.
CNN Images: Adobe Creative Cloud