Highly Accurate Occupational Pneumoconiosis Staging via Dark Channel Prior-Inspired Lesion Area Enhancement.

Abstract

Occupational pneumoconiosis (OP) staging is the core for OP diagnosis. It is essentially an image classification task concerning patients' lung condition by analyzing their chest X-ray. To perform artificial intelligence-assisted OP staging, the chest X-ray film representational learning and classification are commonly adopted, where a convolutional neural network (CNN) has proven to be very efficient. However, unlike commonly encountered image classification tasks, the OP staging relies heavily on the profusion level of opacities, i.e., the OP lesion reflection on the X-ray film. The OP lesions overlap with other tissues in the chest, making the opacities hard to be represented by a standard CNN and thus leading to inaccurate staging results. Inspired by the similarity between OP lesion and haze, i.e., they are both read like dusts in a space, this study proposes a dark channel prior-inspired lesion area enhancement (DCP-LAE)-based OP staging method with high accuracy. Its ideas are twofold: a) enhancing the OP lesion areas with an OP X-ray film restore method inspired by the dark channel prior-based de-hazing method, and b) implementing the multiple feature fusion via a bi-branch network structure to obtain high staging accuracy. Experimental results from real OP cases collected in hospitals demonstrate that the DCP-LAE-based OP staging model achieves an accuracy of 83.8%, surpassing existing state-of-the-art models.