FMDNN: A Fuzzy-guided Multi-granular Deep Neural Network for Histopathological Image Classification

Weiping Ding*, Tianyi Zhou, Jiashuang Huang, Shu Jiang, Tao Hou and Chin-Teng Lin

MOTIVATION

• Histopathological images commonly encompass a variety of morphological features, existing methods neglect the features of histopathological images at diverse granularity levels, limiting feature extraction to a singular scale, which leads to the incomplete capture of intrinsic feature information among cells.

• Feature extraction is often performed at a single granularity, overlooking the multi-granular characteristics of cells.

INNOVATION

• We introduce a multi-granular feature extraction method and employ feature visualization to investigate the semantic scales of features at different granularities, which contributes to a more comprehensive interpretation of multi-granular information embedded in histopathological images.

• We use fuzzy logic to amalgamate key features from images through the computation of universal fuzzy features using three membership functions, which are utilized to globally guide the learning of multi-granular features. Experiments demonstrate the effectiveness of this approach in enhancing classification accuracy during model-guided training while minimizing interference from unrelated features.

• To address information redundancy in multi-granular feature extraction, we employ fuzzy-guided cross-attention to enhance the importance of key features. Continuously guiding features at three granularities with universal fuzzy features reduces interference from irrelevant information, significantly improving classification performance, and achieves good results on five public datasets relative to other algorithms.

  
  

METHOD

Fig. 1. Overview of proposed FMDNN.

As shown in Fig. 1, FMDNN consists of three modules: theMulti-granular Feature Extraction Module conducts feature extraction on the input image at three distinct granularities, theUniversal Fuzzy Feature Module extracts the universal fuzzy features of the image, and theFuzzy-guided Cross-attention Module performs feature fusion through linear transformation and dimension alignment to obtain the final classification result.

• Multi-granular Feature Extraction Moduleemploys a multi-granularity convolution technique to extract features from the original image at various granular levels.

• Universal Fuzzy Feature Module,through the fusion of membership functions tailored to distinct feature points, extracts the universal fuzzy features of the original image. Extracted multi-granular features and universal fuzzy features are separately embedded with patches at different granularities, harmonizing dimensions and reducing subsequent algorithmic complexity.

• Fuzzy-guided Cross-attention Module,which utilizes the universal fuzzy features to guide training at distinct granularities. By employing a multi-granularity classifier as an intermediary, features from different granularities and universal fuzzy features are integrated through fusion learning of feature information and are subsequently back-projected into their respective branches.

EXPERIMENTAL RESULTS

TABLE Ⅰ RESULTS OF COMPARATIVE EXPERIMENTS

Fig. 2. Heatmap and CAM map of label "colonca:" (a) original image; (b) FMDNN fine-, medium-, and coarse-grained feature heatmap and CAM map in order; (c) EfficientNet feature heatmap and CAM map in order; (d) AlexNet feature heatmap and CAM map in order.

Fig. 3. Heatmap and CAM map of label "lungaca:" (a) original image; (b) FMDNN fine-, medium-, and coarse-grained feature heatmap and CAM map in order; (c) EfficientNet feature heatmap and CAM map in order; (d) AlexNet feature heatmap and CAM map in order.

Fig. 4. Comparison of ablation experiments on different datasets: (a) LC; (b) NCT.

Link:https://ieeexplore.ieee.org/abstract/document/10552048