Scholarlink

<p>Colorectal cancer prevention relies on the accurate detection and characterization of precancerous polyps during colonoscopy. However, subtle flat and sessile serrated lesions often exhibit low contrast, ambiguous morphology, and imprecise visual boundaries, leading to high miss rates and uncertain malignancy assessment. We propose FourierBIT2-Net, a unified spectral–temporal deep learning and fuzzy inference framework for real-time colorectal polyp detection, segmentation, and malignancy risk stratification with quantified predictive uncertainty. The pipeline begins by enhancing colonoscopy frames in the YCrCb color space, where the luminance channel is transformed using a fast Fourier transform (FFT) with adaptive spectral masking to highlight subtle frequency-domain cues, while multi-orientation Gabor filtering amplifies faint edges and textures without altering chromatic fidelity. The overall framework explicitly models both epistemic and aleatoric uncertainty. The enhanced frames are then processed by a YOLOv12 backbone equipped with group Fourier-channel attention (GFCA) and spectral-aware multi-scale fusion for frequency-aware feature extraction. This stage jointly predicts bounding boxes, segmentation masks, lesion-type classification, and a preliminary malignancy risk score. To bridge deep representations with clinical interpretability, a neural-guided fuzzy feature extraction module generates interval type-2 fuzzy descriptors of polyp size, texture, and shape, while a hierarchical MedFormer enforces temporal consistency across video sequences. At the decision level, a Bayesian interval type-2 Fuzzy Inference System (BIT2-FIS) integrates interpretable fuzzy rules, Karnik–Mendel type-reduction, and credibility-weighted inference to produce calibrated malignancy risk scores with uncertainty bounds. Experiments on the LDPolypVideo and Kvasir-SEG datasets demonstrate that FourierBIT2-Net achieves 94.80% mAP@0.5 for detection, 93.70% Dice for segmentation, and an AUROC of 0.996 for malignancy risk stratification. Importantly, the proposed framework attains 90.20% sensitivity for diminutive polyps (&lt; 5&#xa0;mm), reducing false negatives by 41.30% compared with baseline methods. These results indicate that FourierBIT2-Net provides accurate, interpretable, and uncertainty-aware decision support, making it well suited for real-time clinical deployment in AI-assisted colonoscopy.</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

FourierBIT2-Net: An Uncertainty-Aware Real-Time Framework for Colorectal Polyp Detection and Cancer Risk Stratification Based on Bayesian Interval Type-2 Fuzzy Logic

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月29日

查看全文
  • Basanta Haobijam,
  • Wei-Che Huang,
  • Duan-You Zheng,
  • Ying-Liang Lu,
  • Cheng-Ming Huang

摘要

Colorectal cancer prevention relies on the accurate detection and characterization of precancerous polyps during colonoscopy. However, subtle flat and sessile serrated lesions often exhibit low contrast, ambiguous morphology, and imprecise visual boundaries, leading to high miss rates and uncertain malignancy assessment. We propose FourierBIT2-Net, a unified spectral–temporal deep learning and fuzzy inference framework for real-time colorectal polyp detection, segmentation, and malignancy risk stratification with quantified predictive uncertainty. The pipeline begins by enhancing colonoscopy frames in the YCrCb color space, where the luminance channel is transformed using a fast Fourier transform (FFT) with adaptive spectral masking to highlight subtle frequency-domain cues, while multi-orientation Gabor filtering amplifies faint edges and textures without altering chromatic fidelity. The overall framework explicitly models both epistemic and aleatoric uncertainty. The enhanced frames are then processed by a YOLOv12 backbone equipped with group Fourier-channel attention (GFCA) and spectral-aware multi-scale fusion for frequency-aware feature extraction. This stage jointly predicts bounding boxes, segmentation masks, lesion-type classification, and a preliminary malignancy risk score. To bridge deep representations with clinical interpretability, a neural-guided fuzzy feature extraction module generates interval type-2 fuzzy descriptors of polyp size, texture, and shape, while a hierarchical MedFormer enforces temporal consistency across video sequences. At the decision level, a Bayesian interval type-2 Fuzzy Inference System (BIT2-FIS) integrates interpretable fuzzy rules, Karnik–Mendel type-reduction, and credibility-weighted inference to produce calibrated malignancy risk scores with uncertainty bounds. Experiments on the LDPolypVideo and Kvasir-SEG datasets demonstrate that FourierBIT2-Net achieves 94.80% mAP@0.5 for detection, 93.70% Dice for segmentation, and an AUROC of 0.996 for malignancy risk stratification. Importantly, the proposed framework attains 90.20% sensitivity for diminutive polyps (< 5 mm), reducing false negatives by 41.30% compared with baseline methods. These results indicate that FourierBIT2-Net provides accurate, interpretable, and uncertainty-aware decision support, making it well suited for real-time clinical deployment in AI-assisted colonoscopy.