DC-MSSFF Net: Dule-channel multi-scale spatial-spectral feature fusion network for cholangiocarcinoma pathology high-resolution hyperspectral image segmentation

• Published in: Computer methods and programs in biomedicine Vol. 269; p. 108905
• Main Authors: Liang, Meiyan, Xi, Zelin, Li, Bo, Wang, Lin
• Format: Journal Article
• Language: English
• Published: Ireland Elsevier B.V 01.09.2025
• Subjects: Algorithms; Bile Duct Neoplasms - diagnostic imaging; Bile Duct Neoplasms - pathology; Cholangiocarcinoma; Cholangiocarcinoma - diagnostic imaging; Cholangiocarcinoma - pathology; Ensemble learning; Graph convolutional network; Graph-within-graph; Humans; Hyperspectral image; Hyperspectral Imaging - methods; Image Processing, Computer-Assisted - methods; Neural Networks, Computer; Graph convolutional network; Graph-within-graph; Ensemble learning; Hyperspectral image; Cholangiocarcinoma
• ISSN: 0169-2607; 1872-7565; 1872-7565
• DOI: 10.1016/j.cmpb.2025.108905

High-precision segmentation of pathological images is a challenging task in the field of medical image processing. Hyperspectral microscopic imaging offers a distinct advantage in histopathological image segmentation due to its abundance of spectral and spatial data. Here, a Dule-Channel Multi-Scale Spatial-Spectral Feature Fusion Network (DC-MSSFF Net) is proposed for semantic segmentation of cholangiocarcinoma hyperspectral images (HSI). The DC-MSSFF Net is composed of two parallel channels, graph-within-graph (GwG) and multi-scale CNN. The GwG can greatly reduce the computational burden while establishing the spatial context relationship of the HSI image. The multi-scale CNN channel is able to fine-tune the segmented edges of the HSI images at the pixel-level based on hyperspectral information in the depth dimension. Afterwards, the segmentation results are achieved by fusing the features from the two channels. Furthermore, an ensemble-based framework is applied to further improve the performance of the model. The image segmentation evaluation indexes such as dice similarity coefficient (Dice) of the Cholangiocarcinoma HSI data can reach 70.47, which is much higher than the SOTA method and RGB-based image segmentation methods. The superior performance of the DC-MSSFF network pioneers the inductive learning task of deep frameworks for semantic segmentation of high-resolution hyperspectral image (HR-HSI). •Inductive GCN for HR-HSI Segmentation: Drawing inspiration from GraphSAGE, we successfully integrate neighbor sampling and feature aggregation into our Graph-within-Graph backbone. This integration not only significantly enhances the inductive capability of the DC-MSSFF Net but also improves computational efficiency.•Hierarchical Contextual Feature Representation: We introduce a hierarchical Graph-within-Graph dual-topology representation that embeds subgraphs into a global graph structure, effectively integrating local and global HSI features. This approach balances the need for context-aware feature representation with the computational burden, offering a robust framework for efficient processing.•Fine-Tuning and Optimization: The multi-scale CNN branch leverages multi-scale spectral-spatial features to further refine segmentation accuracy. It fine-tunes the model to mitigate the impact of heterogeneous pixels introduced during the preprocessing stage and captures subtle feature variations, thereby generating more accurate superpixel boundaries.•Ensemble Learning: By employing ensembles of three base learners configured with different hyperparameters, we further enhance the accuracy and reliability of HSI segmentation results through a majority voting mechanism. This ensemble strategy not only improves segmentation performance but also strengthens the model’s generalization capability, which is crucial for practical applications in pathology.