Liang Xiong ¹ Liu Xin ^2* Qin X. lin ^1* Weiling Li ^2*

• ¹ Chengdu Institute of Computer Application, Chinese Academy of Sciences (CAS), Chengdu, China
• ² Dongguan University of Technology, Dongguan, Guangdong, China

Accurate pneumoconiosis staging is the key to early intervention and treatment planning for pneumoconiosis patient. The staging result heavily depends on the profusion level of small opacities, which disperse within the entire lung field and manifest as kinds of textures. Conventional convolutional neural network via learning characteristics of one whole large object has made significant success in image classification, object recognition and many other applications, but is not ideal for classifying fine-grained medical images due to the need for global orderless feature representation, and thereby leads to inaccurate pneumoconiosis staging result. In this paper, we construct a deep texture encoding scheme with suppression strategy to learn the global orderless information of pneumoconiosis lesions and suppress the salient regions such as ribs and clavicles in the lung field. In order to leverage the ordinal information among profusion levels of opacities, we use an ordinal label distribution for the representation learning. In addition, we adopt supervised contrastive learning to obtain a discriminative feature space for downstream classification. Finally, conforming to the standard, we evaluate the profusion level of opacities of each subregion rather than the whole chest X-ray image. Experimental results on the pneumoconiosis dataset demonstrate the superior performance of the proposed method.