Domain adaptation and weight initialization of neural networks for diagnosing interstitial lung diseases

Abstract

Interstitial lung diseases (ILDs) are adverse disorders, damaging the lung tissues, thus making timely diagnosis imperative. To counter the scarcity of publicly available high‐resolution computed tomography (HRCT) data, architecture employing HRCT and x‐ray data have been proposed for diagnosing ILDs. A model is first trained to diagnose three ILDs and a healthy lung from x‐ray data and then with a small amount of HRCT data of the same four classes. We introduce an EfficientNet+AlexNet model and a custom deep convolutional neural network and compare their performances with different weight initialization at the start of x‐ray image training. The EfficientNet+AlexNet model initialized with ImageNet weights performed best on the x‐ray dataset and was later used for domain adaptation. We also compare our model's performance to a single pre‐trained EfficientNetB0 model (trained on the same HRCT data). Our model gave an accuracy of 97.4% on the testing dataset, a notable increase in the performance compared to the recently proposed methodologies. The experimental results and graphical plots depict the superiority of domain adaptation and the potentials of weights that are a starting point for training the models.