CFT-Triplet: integrating contextual, frequency and textural features for fine-grained pulmonary fissure detection
4 affiliations
- 1Shandong Public Health Clinical Center , ,
- 2Linyi People's Hospital , ,
- 3Broncus Medical Inc. , ,
- 4Anhui Chest Hospital , ,
Abstract
Accurate detection and integrity assessment of pulmonary fissures are essential for the effective treatment of emphysema and chronic obstructive pulmonary disease (COPD). However, clinical CT scans often present fissures with weak contrasts and severe deformations. Standard deep learning approaches face two fundamental limitations in this task: First, they exhibit an inherent “spectral bias,” preferentially learning low-frequency shapes while suppressing high-frequency boundary details (i.e., the fissures). Second, existing Transformer-based global modeling incurs quadratic computational costs, limiting the efficient capture of long-range anatomical continuity. To address these challenges, we propose CFT-Triplet, a framework that integrates Contextual, Frequency, and Textural features for robust fissure detection. The core of our framework is the CFT-Triplet Block, an architectural unit designed to extract and fuse these three complementary features within a unified module. Within each Triplet Block, we design a hybrid topology: (1) To model long-range anatomical dependencies, we incorporate Bidirectional Mamba2, which captures global Contextual information with linear complexity; (2) To specifically compensate for spectral loss, these contextual features are serially refined by a dedicated Frequency module, sharpening high-frequency boundary components; (3) Simultaneously, a parallel CNN-based Textural branch is utilized to extract local spatial patterns. This triplet design ensures that every network layer possesses both the global receptive field and the sensitivity to local edges. The multi-view features are then fused via a Gate-Net module to maximize complementarity. Extensive experiments on the Shandong Provincial Chest Hospital (SPCH) dataset—where 400 high-quality CT scans were selected from 500 collected cases after rigorous data cleaning (partitioned into 320 training, 40 validation, and 40 testing sets at the patient level to prevent data leakage)—demonstrate that CFT-Triplet outperforms state-of-the-art methods on our private SPCH dataset.
Related
- Integrating multimodal data for precise subtyping and prognostication in patients with pulmonary mucinous adenocarcinoma
- Nomogram models integrating multiregion radiomics, quantitative computed tomography, and clinical indicators for distinguishing connective tissue disease-associated interstitial lung disease from idiopathic pulmonary fibrosis and improving the prediction of gender-age-physiology stage
- Multi-scale contextual modeling and fine-grained adaptive fusion for real-time surface defect detection
- Site‐Specific Characterization and Quantification of Facial Corticosteroid Addiction Dermatitis: A Multimodal Nomogram Integrating High‐Frequency Ultrasonography and Transepidermal Water Loss
- Identifying a Cardio‐Hepatic Phenotype by Integrating the Albumin‐Bilirubin Score Into Risk Assessment in Pulmonary Arterial Hypertension
BibTeX
@article{xu2026cfttriplet,
title = {{CFT-Triplet: integrating contextual, frequency and textural features for fine-grained pulmonary fissure detection}},
author = {Li Xu and Yu Jiang and Rong Tao and Nanyu Li and Riqing Chen and Chenhui Su and Guodong Li and Fei Tang},
journal = {Frontiers in artificial intelligence},
year = {2026},
month = oct,
doi = {10.3389/frai.2026.1817367},
url = {https://europepmc.org/article/PMC/PMC13727603},
note = {CC BY},
}
RIS
TY - JOUR
TI - CFT-Triplet: integrating contextual, frequency and textural features for fine-grained pulmonary fissure detection
AU - Li Xu
AU - Yu Jiang
AU - Rong Tao
AU - Nanyu Li
AU - Riqing Chen
AU - Chenhui Su
AU - Guodong Li
AU - Fei Tang
PY - 2026
DA - 2026/10/20
JO - Frontiers in artificial intelligence
DO - 10.3389/frai.2026.1817367
UR - https://europepmc.org/article/PMC/PMC13727603
AB - Accurate detection and integrity assessment of pulmonary fissures are essential for the effective treatment of emphysema and chronic obstructive pulmonary disease (COPD). However, clinical CT scans often present fissures with weak contrasts and severe deformations. Standard deep learning approaches face two fundamental limitations in this task: First, they exhibit an inherent “spectral bias,” preferentially learning low-frequency shapes while suppressing high-frequency boundary details (i.e., the fissures). Second, existing Transformer-based global modeling incurs quadratic computational costs, limiting the efficient capture of long-range anatomical continuity. To address these challenges, we propose CFT-Triplet, a framework that integrates Contextual, Frequency, and Textural features for robust fissure detection. The core of our framework is the CFT-Triplet Block, an architectural unit designed to extract and fuse these three complementary features within a unified module. Within each Triplet Block, we design a hybrid topology: (1) To model long-range anatomical dependencies, we incorporate Bidirectional Mamba2, which captures global Contextual information with linear complexity; (2) To specifically compensate for spectral loss, these contextual features are serially refined by a dedicated Frequency module, sharpening high-frequency boundary components; (3) Simultaneously, a parallel CNN-based Textural branch is utilized to extract local spatial patterns. This triplet design ensures that every network layer possesses both the global receptive field and the sensitivity to local edges. The multi-view features are then fused via a Gate-Net module to maximize complementarity. Extensive experiments on the Shandong Provincial Chest Hospital (SPCH) dataset—where 400 high-quality CT scans were selected from 500 collected cases after rigorous data cleaning (partitioned into 320 training, 40 validation, and 40 testing sets at the patient level to prevent data leakage)—demonstrate that CFT-Triplet outperforms state-of-the-art methods on our private SPCH dataset.
SN - 2624-8212
ER -