This repo comes from a final project of Fundamental of Control and Intelligence System at Institut Teknologi Bandung by: Robertsen Putra Sugianto
COVID-19 is a pandemic with a very fast spreading time and symptoms common with other coronavirus diseases. The key to this pandemic is early and accurate detection so that its spread can be suppressed. One method detection of COVID-19 besides PCR and rapid test is the detection of COVID-19 in the image CXR.
- The existence of false negative detection by PCR at the beginning of infection CXR (Chest X-ray) images can be complementary to the initial PCR results so as to reduce false negatives as much as possible.
- The PCR process is more accurate, but takes a long time and not many labs in Indonesia are capable. Radiology images tend to be fast and can be done anywhere.
COVID-19 is a viral infection that can experienced by anyone. X-rays has the advantage of being cheap and price fast; therefore, X-rays are easier accessed by health care providers who work in a small area and / or isolated. This model is a prototype system and not for medical use and cannot be used for diagnosis yet.
The number of datasets used in this task is as many as 266 pieces of CXR images, with information:
- 160 thorax data for COVID-19 patients
- 106 normal thorax data
Normal x-ray images
Covid x-ray images
Deep Learning method is used to do classification of Pneumonia (Covid-19) on radiological images by looking for signs of appearance infiltrates / opacities / consolidation.
Transfer learning method is used using the pretrained model VGG16 have been trained using the Imagenet dataset. VGG-16 is a CNN architecture, used for feature extraction in the image. Pretrained VGG-16 as the base model, then add a new model(trainable) on the final layer of VGG-16.
The covid19 Plot Model used in this project. There is a change in the number of neurons in the dense_8 layer from 64 to 128
By using this dataset, training accuracy is obtained around 90%. Sklearn library with GridSearch class is used to sweep some of the optimum values of hyperparameters. This sweeping starts with coarse tuning and fine tuning. The following results were obtained:
- Learning Rate (LR) = 0.008
- Batch Size (BS) = 7
- Epoch = 12
- Optimizer = Adam
Grid Search supports the K-fold cross validation feature which is used to prevent overfitting. Adam optimizer is used because it includes RMSprop and Adagrad as well.
Accuracy = 0.9815 Precision = TP / (TP + FP) = 0.97 Sensitivity = TP / (TP + FN) = 1,000 Specificity = TN / (TN + FP) = 0.9545