Hyperspectral pathology image classification using dimension-driven multi-path attention residual network

Abstract

Hyperspectral imaging technology (HSI) can capture pathological tissue’s spatial and spectral information simultaneously, with wide coverage and high accuracy characteristics, and is widely used in biomedical imaging. As an image-spectrum merging technology, HSI can obtain more practical information in disease diagnosis, which is helpful for pathological analysis. Focusing on the characteristics of scattered distribution of pathological areas, and combined with the advantages of HSI technology, a dimension-driven multi-path attention residual network (DDMARN) is proposed to pixel-level classification for membranous nephropathy (MN). To make full use of the space-spectrum information of hyperspectral data, dimension-driven multi-path attention residual block (DDMARB) is developed to effectively obtain the multi-scale features and differently treats these features containing different amounts of information through the channel attention (CA) mechanism, which makes the data depth features better expressed. The experimental results demonstrate that the proposed DDMARN performs very competitively in the PMN and HBV-MN classification tasks. The average of OA (%) ± standard deviation (%), AA (%) ± standard deviation (%), and Kappa coefficient ± standard deviation of 10 experiments results are 96.23 ± 0.17, 90.24 ± 0.18, and 0.8654 ± 0.0056, respectively, the optimal values in the comparison algorithm, and the model parameter is only 700K.