Abstract
Early detection of gastric lesions is critical for patient survival, yet automated segmentation remains challenging due to noise and visual similarities between healthy and pathological tissues. This paper presents Deep DuS-KFCM, a hybrid "Neuro-Fuzzy" framework that integrates Dual Spatial Kernelized Constrained Fuzzy C-Means (KFCM) with DeepLabv3+. By combining statistical texture analysis with deep learning refinement, the model achieves up to 98% accuracy in gastric bleeding detection, providing a robust tool for clinical diagnosis.
Introduction
Gastric cancer remains a significant global health burden. Early-stage detection is the primary determinant of clinical outcomes, with five-year survival rates exceeding 96% when caught early, compared to less than 40% for late-stage diagnoses. Endoscopic imaging is the standard diagnostic tool; however, the manual identification of bleeding regions is labor-intensive and subjective. Automated systems often struggle with the subtle intensity variations and noise inherent in endoscopic imagery. To address these limitations, we propose a coarse-to-fine segmentation strategy that leverages both traditional fuzzy logic and modern deep learning.
Methodology: The Deep DuS-KFCM Framework
The proposed Deep DuS-KFCM model operates in two distinct stages to ensure both sensitivity and structural precision.
Stage 1: Coarse Segmentation via Fuzzy Logic
The initial stage employs a modified Fuzzy C-Means (FCM) algorithm. Unlike standard approaches that rely solely on color, this "Dual Spatial" method extracts 22 statistical texture features using a Gray Level Co-occurrence Matrix (GLCM) alongside RGB color histograms. These features capture essential characteristics such as energy, contrast, and entropy. A Correlation-based Feature Selection (CFS) mechanism is then applied to isolate the most discriminative features, producing an initial coarse map of the lesion.
Stage 2: Fine Refinement via Deep Learning
The coarse output is subsequently refined using a DeepLabv3+ architecture with a ResNet50 backbone. This deep learning component functions as a spatial "cleaner," utilizing hierarchical feature extraction to sharpen boundaries and eliminate spurious noise regions identified in the first stage. This synergy allows the model to benefit from the mathematical stability of fuzzy clustering and the contextual intelligence of convolutional neural networks (CNNs).
Experimental Setup and Datasets
The model was validated using several benchmark datasets:
- Kvasir-SEG: A standard dataset for gastrointestinal polyp segmentation.
- Gastric Bleeding (GB) Dataset: Focused on various stages of active bleeding.
- Gastric Red Spots (GRS) Dataset: Used to evaluate sensitivity to subtle, early-stage symptoms. Ground truth annotations were provided by expert gastroenterologists to ensure benchmark reliability.
Results and Discussion
The Deep DuS-KFCM model demonstrated superior performance compared to traditional methods such as standard Fuzzy K-Means (FKM) and Gaussian Mixture Models (GMM). Key performance metrics include:
- Accuracy: The model achieved a benchmark accuracy of 87.95% on the difficult GRS dataset and up to 98% on general gastric endoscopy images.
- Specificity: A high specificity of 96.33% was recorded, highlighting the model's ability to minimize false positives in healthy tissue.
- Qualitative Alignment: Visual comparisons showed that the automated segmentation boundaries closely matched expert-annotated ground truths, effectively delineating complex lesion margins.
Conclusion and Future Work
The Deep DuS-KFCM framework represents a significant advancement in automated gastric lesion detection. By integrating texture-aware fuzzy clustering with deep learning, the system overcomes the limitations of single-modality models. Future research will focus on expanding the model's scope to detect metastatic cancers of non-gastric origin and exploring the application of this "Neuro-Fuzzy" logic to high-dimensional genomic data clustering.
Summary updated