A multi-granular hybrid dual-branch convolutional architecture for texture-aware histopathological tissue classification

In computational pathology, accurate multi-class tissue classification facilitates the timely detection of early stages of cancer, propitious to meticulous diagnostic chains. However, the majority of the histopathological deep learning (DL) frameworks are confined by global convolutional capabilities, inadequate sensitivity to texture patterns, and generalizability constrained from a non-binary space. To address this limitation, we propose MGH-DBNet, a multi-granular hybrid dual branch convolution neural network designed specifically for eight-class gastric histopathology classification using the GCHTID dataset, which consists of 31,096 H&E-stained non-overlapping images. Texture cues are essential, as micro-structural patterns often drive accurate tissue differentiation. This multi-granular hybrid convolutional architecture includes a global feature-extraction branch built on a convolutional backbone and a lightweight local-texture encoder of Local Binary Patterns-Gray Level Co-occurrence Matrix branch designed to capture fine-grained texture patterns. Unlike simply deepening the CNN, which increases parameters without guaranteeing sensitivity to micro-textures, LBP-GLCM explicitly encodes boundary irregularities and spatial intensity dependencies, enabling MGH-DBNet to capture discriminative micro-structural patterns that deep CNNs typically miss. These two streams are integrated using an attention mechanism that adaptively emphasizes the most informative complementary cues. A comparative evaluation of contemporary Convolution Neural Networks, like DenseNet121 of similar configurations, revealed that the proposed network outperformed all other models, attaining an AUC score of 0.954. Furthermore, Grad-CAM, a heat-map-based visualization method, renders the model engaged in clinically pertinent tissue regions for improved explainability. In addition, these results maintain stability across different preprocessing settings, backbone selection, and fusion strategies during ablation studies, indicating high robustness and generalizability of the proposed MGH-DBNet framework. Pursuant to an exhaustive survey, the authors opine that the proffered solution intended for multi-class histopathology entails a distinct computationally efficient architecture that endorses quality assessment and decision support for real-world applications.