Truncating a densely connected convolutional neural network with partial layer freezing and feature fusion for diagnosing COVID-19 from chest X-rays

Montalbo, Francis Jesmar P.

Deep learning and computer vision revolutionized a new method to automate medical image diagnosis. However, to achieve reliable and state-of-the-art performance, vision-based models require high computing costs and robust datasets. Moreover, even with the conventional training methods, large vision-based models still involve lengthy epochs and costly disk consumptions that can entail difficulty during deployment due to the absence of high-end infrastructures. Therefore, this method modified the training approach on a vision-based model through layer truncation, partial layer freezing, and feature fusion. The proposed method was employed on a Densely Connected Convolutional Neural Network (CNN), the DenseNet model, to diagnose whether a Chest X-Ray (CXR) is well, has Pneumonia, or has COVID-19. From the results, the performance to parameter size ratio highlighted this method's effectiveness to train a DenseNet model with fewer parameters compared to traditionally trained state-of-the-art Deep CNN (DCNN) models, yet yield promising results. • This novel method significantly reduced the model's parameter size without sacrificing much of its classification performance. • The proposed method had better performance against some state-of-the-art Deep Convolutional Neural Network (DCNN) models that diagnosed samples of CXRs with COVID-19. • The proposed method delivered a conveniently scalable, reproducible, and deployable DCNN model for most low-end devices.