Scholarlink

Background <p>Otoscopic image analysis plays a critical role in diagnosing ear pathologies; however, existing artificial intelligence approaches often face challenges related to interpretability, robustness under acquisition variability, and reliable feature integration.</p> Objective <p>This study proposes a reliability-aware multi-feature fusion framework for ear disease classification by integrating complementary visual representations texture, color, and shape derived from a single otoscopic imaging modality.</p> Methods <p>Texture features are extracted using a Vision Transformer (ViT) with confidence-weighted patch enhancement, while color information is captured using CIE Lab histograms and statistical descriptors, and shape features are derived through clinically guided contour analysis following tympanic membrane segmentation. These feature representations are adaptively integrated using a Mamdani fuzzy inference system based on feature branch-specific reliability scores. Performance is evaluated using accuracy, macro-F1 score, AUROC, and negative log-likelihood (NLL), along with class-wise sensitivity and specificity. Robustness is assessed under variations in illumination, color, and blur.</p> Results <p>The proposed framework achieves 97.5% accuracy, 0.96 macro-F1 score, 0.98 AUROC, and 0.12 NLL, outperforming individual feature-based baselines (Texture: 92.0%, Color: 94.2%, Shape: 95.0%) and conventional late averaging (96.0%, 0.94 F1, NLL 0.14). Consistent performance is observed across all disease classes, with sensitivity and specificity values exceeding 0.95. Robustness analysis demonstrates improved performance under illumination (90.8% → 95.0%), hue/saturation shifts (91.2% → 95.5%), and blur (89.0% → 94.0%). Interpretability is enhanced through confidence-weighted patch maps, fused attention overlays, and feature branch attribution.</p> Conclusion <p>The proposed framework demonstrates that reliability-aware multi-feature fusion within a single imaging modality can achieve robust, interpretable, and clinically meaningful ear disease classification, offering a practical alternative to complex multimodal systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A reliability-aware multi-feature fusion framework using fuzzy logic for interpretable ear disease classification from otoscopic images

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月27日

查看全文
  • Eshika Jain,
  • Vinay Kukreja,
  • Pratham Kaushik,
  • Anuj Kumar Jain,
  • Malvinder Singh Bali

摘要

Background

Otoscopic image analysis plays a critical role in diagnosing ear pathologies; however, existing artificial intelligence approaches often face challenges related to interpretability, robustness under acquisition variability, and reliable feature integration.

Objective

This study proposes a reliability-aware multi-feature fusion framework for ear disease classification by integrating complementary visual representations texture, color, and shape derived from a single otoscopic imaging modality.

Methods

Texture features are extracted using a Vision Transformer (ViT) with confidence-weighted patch enhancement, while color information is captured using CIE Lab histograms and statistical descriptors, and shape features are derived through clinically guided contour analysis following tympanic membrane segmentation. These feature representations are adaptively integrated using a Mamdani fuzzy inference system based on feature branch-specific reliability scores. Performance is evaluated using accuracy, macro-F1 score, AUROC, and negative log-likelihood (NLL), along with class-wise sensitivity and specificity. Robustness is assessed under variations in illumination, color, and blur.

Results

The proposed framework achieves 97.5% accuracy, 0.96 macro-F1 score, 0.98 AUROC, and 0.12 NLL, outperforming individual feature-based baselines (Texture: 92.0%, Color: 94.2%, Shape: 95.0%) and conventional late averaging (96.0%, 0.94 F1, NLL 0.14). Consistent performance is observed across all disease classes, with sensitivity and specificity values exceeding 0.95. Robustness analysis demonstrates improved performance under illumination (90.8% → 95.0%), hue/saturation shifts (91.2% → 95.5%), and blur (89.0% → 94.0%). Interpretability is enhanced through confidence-weighted patch maps, fused attention overlays, and feature branch attribution.

Conclusion

The proposed framework demonstrates that reliability-aware multi-feature fusion within a single imaging modality can achieve robust, interpretable, and clinically meaningful ear disease classification, offering a practical alternative to complex multimodal systems.