ChestAidNet: leveraging lightweight CNN to detect chest diseases from radiology images

Abstract

Autonomous methods to determine chest diseases in X-ray images deal with the problem of computational expense and accuracy in the diagnosis. Although deep learning models are highly accurate, their requirements are very high, which restricts their implementation in a clinical setting especially in resource-restricted environments. In this study, ChestAidNet, a convolutional neural network with lightweight attention-based characteristics to classify chest diseases of NIH ChestX-ray14 (112,120 images, 15 disease classes) is introduced. ChestAidNet is designed with depthwise separable convolutions, channel and spatial attention, and residual connections that can deliver computational efficiency without sacrificing diagnostic performance. The models were trained on 72% of the dataset while the validation and testing were done on 8% and 20% of the dataset and various data augmentation techniques were used. ChestAidNet uses 5.09M parameters (a 78-percent reduction over ResNet-50 (23.5M) and 27-percent reduction over DenseNet-121 (7.0M)) and has 94% accuracy which is equivalent or better diagnostic accuracy compared to them. Single images are processed in 40.70 ms with throughput of 251.76 images/second, making the model capable of diagnosing in real-time. The comparison with ResNet-50, DenseNet-121, InceptionV3 and Xception proves that ChestAidNet is characterized by the most optimal tradeoff between accuracy and computational cost. Grad-CAM visualization verifies clinically significant feature responses, concentrating on pathologically meaningful areas. ChestAidNet can be deployed on mobile and in resource-constrained clinical environments due to its compact size (58.57 MB) and fast inference, which fills important gaps in diagnostic technology available to the chest disease detection field.