Data Augmentation and pre-processing

For Unet Lung Segmentation training 941 CXRs and corresponding manual lung mask pairs from sourced from SCR database, Shenzhen and Montgomery County of NIH database. Further augmentation was performed to increase the data points for the training. After the augmentation, additional pre-processing was performed on the CXR masks, thresholding the diluted pixel back to 0 values (black). Further, the CXRs used for the segmentation model were pre-processed using CLAHE followed by gaussian blurring to mimic the COVID positive CXRs, thus enabling segmentation model to learn on abnormal CXRs.

For Classification trainings, COVID-19 Radiography database from Kaggle was used consisting of 3616 Covid-19 infected CXRs and 10192 Normal labeled CXRs. As the class sizes of our data set was unbalanced which could lead to biased learning. Hence, class balancing was done to overcome overfitting by augmenting the training data, which was achieved by adding random variations in the following attributes of the images: rotations, translational shifts, zooming, shear shifts, horizontal flips, and brightness intensity changes.

Unet lung segmentation and lung RoI extraction

For training our classification model we extracted the Lung region of interest (ROI) from the CXRs to limit our model to learn only from the Lung region of the CXRs and avoid any kind of mis-learning. For this we trained a Unet lung segmentation model using the Adam optimizer on manual lung masks and achieved test accuracy of 98.3% and IoU test score of 0.925. and image cropping. We then performed contour detection over the lung masks. A union of the fit bounding boxes around the contours was cropped, providing us ROI approximation of the thoracic cavity consisting of the lung lobes. These images were used to train our DL classification models only the relevant features, thereby aiding reliable decision-making.

Transfer learning-based Classification training

We have trained 5 different model on ImageNet pre-trained CNN for architectures: VGG-16, ResNet50, DenseNet201 (DNet201), InceptionV3 (InceptV3), and Xception in our study and a final ensemble model was developed out of the whole training. The pre-trained models were fine-tuned by truncating the last convolution layer, and classifier layers are further replaced by custom build layers. We added an average pooling layer to reduce the spatial size of representation generated by previous kernels after convolution. Batch normalization has also been included in the architecture of all pre-trained models except for VGG-16. After the convolutional and pooling layers, a classification network is placed at the end of the model consisting of fully connected layers (FCLs). We have used FCL with 256 neurons with ReLU activation function followed by dropout.

Ensemble model

Ensemble techniques use a combination of learning algorithms to optimize better predictive performance. They typically reduce overfitting in models and make the model more generalizable, making it unlikely to be influenced by small changes in the training data. We studied different ensemble techniques for COVID-19 case detection by combining predictions from multiple deep learning models via various ensemble strategies such as hard voting, weighted averaged voting, SVM, XG-boost and stacked generalization

AI explainability using Gradcam

While using DL models in medical image prediction problems, it’s important to understand their learned behavior, to explain the pre-dictions from these black-box models, and understand the clinical decision making. In our proposed work, we have applied a gradient-based approach called Grad-CAM, which measures the gradients of features maps in the final convolution layer on a CNN model for a target image, highlighting the critical regions that are class-discriminating saliency maps.