MSGM: an Advanced Deep Multi-Size Guiding Matching Network for Whole Slide Histopathology Images Addressing Staining Variation and Low Visibility Challenges
Xiaoxiao Li,Zhengxiong Li,Taobo Hu,Mengping Long,Xiao Ma,Jin Huang,Yiqiang Liu,Yaxiaer Yalikun,Sheng Liu,Du Wang,Jianghua Wu,Liye Mei,Cheng Lei
DOI: https://doi.org/10.1109/jbhi.2024.3417937
2024-01-01
Abstract:Matching whole slide histopathology images to provide comprehensive information on homologous tissues is beneficial for cancer diagnosis. However, the challenge arises with the Giga-pixel whole slide images (WSIs) when aiming for high-accuracy matching. Learning-based methods are difficult to generalize well with large-size WSIs, necessitating the integration of traditional matching methods to enhance accuracy as the size increases. In this paper, we propose a multi-size guiding matching method applicable high-accuracy requirements. Specifically, we design learning multiscale texture to train deep descriptors, called TDescNet, that trains 64 x 64 x 256 and 256 x 256 x 128 size convolution layer as C64 and C256 descriptors to overcome staining variation and low visibility challenges. Furthermore, we develop the 3D-ring descriptor using sparse keypoints to support the description of large-size WSIs. Finally, we employ C64, C256, and 3D-ring descriptors to progressively guide refined local matching, utilizing geometric consistency to identify correct matching results. Experiments show that when matching WSIs of size 4096 x 4096 pixels, our average matching error is 123.48 mu m and the success rate is 93.02$\%$ in 43 cases. Notably, our method achieves an average improvement of 65.52 mu m in matching accuracy compared to recent state-of-the-art methods, with enhancements ranging from 36.27 mu m to 131.66 mu m. Therefore, we achieve high-fidelity whole-slice image matching, and overcome staining variation and low visibility challenges, enabling assistance in comprehensive cancer diagnosis through matched WSIs.