Original Source Here
We split the data-set into three sets — train, validation and test sets. We used data augmentation like rotate, shearing, zooming, horizontal and vertical flipping, contrast, shiftRGB, shiftHSV, grayscale, equalize, resize, multiply and brightness change to increase the dataset size to almost double the original dataset size. This operation helps the model prevent overfitting thus making it more generalizable to unseen images in the test set.
In between the convolutional and max pooling layers for each block, we used 30% dropouts to reduce overfitting. Activation function used was Relu throughout except for the last layer where it was Sigmoid as this is a binary classification problem. We used both Adam as the optimizer and binary cross-entropy as the loss function. We tried with pre trained models like Inception v3, InceptionResNet v2 and ResNet 152 by fine tuning the last layers of the network. We used two dense layers with 64 neurons and 2 neurons respectively.
We trained the model for 20 epochs with a batch size value of 16 by doing a grid search for the optimal hyper-parameters values like learning rate, batch size, optimizer and pre-trained weights. We defined the optimal as the maximum value achieved for F1 score which is the weighted average of precision and recall. Since the F1 value takes both false negative and false positive and penalizes them in a single term, hence it would be the most accurate measure of the classifier.
We used model checkpoint while training the neural network and early stopping when the validation loss started increasing. These architectural and training details are presented in the next subsection. We used ModelCheckpoint. Often many iterations are required, when training requires a lot of time to achieve a good result. In this case, it is better to save a copy of the best performing model only when an epoch that improves the metrics ends.
We also used EarlyStopping. Sometimes, during training we can notice that the generalization gap (i.e. the difference between training and validation error) starts to increase, instead of decreasing. This is a symptom of overfitting that can be solved in many ways. (reducing model capacity, increasing training data, data augmentation, regularization, dropout, etc). Often a practical and efficient solution is to stop training when the generalization gap is getting worse. Fig 28 shows early stopping.
The feature maps help in explaining what the model is learning at every layer. As the depth increases, the model is able to learn more spatial information. In other words, the neural networks go on from learning edges and blobs in the first layer to complete objects in the last layers.
Visualization of feature maps is important because this makes hyperparameter tuning easier, since when we see an error made by the neural network then we can know what is making the algorithms going wrong. The functionality and expected behaviour of the networks can be explained, especially to non-technical stakeholders who would often demand explanation before accepting the results. Also we can further extend and improve the overall design of our models since we’d have knowledge of the current design using the learned filters and we can improve that for the future models.
By visualising the learned weights we can get some idea as to how well our network has learned. For example, if we see a lot of zeros then we’ll know we have many dead filters that aren’t doing much for our network which means we need to do some pruning for model compression.
The loss vs epochs and accuracy vs epochs is plotted in figure below. The loss has converged on both the training and the validation set in 60 iterations. The accuracy on the training set is 100% and on validation set is 95%. Since the model was not shown the images on the validation set while training, the accuracy is quite good.
We also used ROC-AUC curves to evaluate the classifier. The 45 degree line is the random line, where the area under the curve or AUC is 0.5 . The further the curve from this line, the higher the ROC-AUC and better the model is at classifying. The highest a model can get is an AUC value of 1, where the curve forms a right angled triangle. This denotes that the classifier is perfect. The ROC curve can also help in debugging the model misclassifications. The ROC- AUC curve of the binary classifier is shown in figure below:
In this paper we demonstrated how to classify positive and negative coronavirus images from a collection of X-ray images. We compared various transfer learning architectures like ResNet50, InceptionV3 and InceptionResNetV2 and found ResNet50 to give the best overall results. For evaluation we used metrics including precision, recall, F1 score and ROC-AUC. Finally we made a comparative analysis using various transfer learning architectures, learning rates, batch size and optimizers. By using grid search hyperparameter optimization, the best results were obtained with a learning rate value of 0.0001, batch size value of 16 and ADAM as the optimizer. Due to shortage of radiologists in the world, it is often the case that coronavirus gets unnoticed. Also many times mistakes are made by radiologists themselves, hence it is a good idea to automate the process for better and more efficient diagnosis.
B. Ghoshal and A. Tucker. Estimating uncertainty and interpretability in deep learning for coronavirus (covid-19) detection. arXiv preprint arXiv:2003.10769, 2020.
O. Gozes, M. Frid-Adar, N. Sagie, H. Zhang, W. Ji, and H. Greenspan. Coronavirus detection and analysis on chest ct with deep learning. arXiv preprint arXiv:2004.02640, 2020. K. He, X. Zhang, S. Ren, and J. Sun.
Identity mappings in deep residual networks. In European conference on computer vision, pages 630–645. Springer, 2016. G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger.
Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700–4708, 2017. D. S. Kermany, M. Goldbaum, W. Cai, C. C. Valentim, H. Liang, S. L. Baxter, A. McKeown, G. Yang, X. Wu, F. Yan, et al.
Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5):1122–1131, 2018. D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25:1097–1105, 2012. P. Rajpurkar, J. Irvin, K. Zhu, B.
Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. Langlotz, K. Shpanskaya, et al. Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225, 2017.
Before You Go
Trending AI/ML Article Identified & Digested via Granola by Ramsey Elbasheer; a Machine-Driven RSS Bot