YOLOv11-MFF: A multi-scale frequency-adaptive fusion network for enhanced CXR anomaly detection

Chest X-ray (CXR) represents one of the most widely utilized clinical diagnostic tools for thoracic diseases. Nevertheless, computer-aided diagnosis based on chest radiographs still faces considerable challenges in anomaly detection. Certain lesions in CXRs exhibit subtle radiographic characteristic...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:PloS one Ročník 20; číslo 10; s. e0334283
Hlavní autori: Guan, Li, Zhang, Ruting, Zhao, Yi
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: United States Public Library of Science 01.10.2025
Public Library of Science (PLoS)
Predmet:
Algorithms
Anomalies
Background noise
Chest
Classification
Datasets
Deep learning
Disease
Fourier transforms
Humans
Lesions
Localization
Modules
Neural networks
Neural Networks, Computer
Pneumonia
Radiographic Image Interpretation, Computer-Assisted - methods
Radiography, Thoracic - methods
Semantics
Thorax
ISSN:1932-6203, 1932-6203
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Chest X-ray (CXR) represents one of the most widely utilized clinical diagnostic tools for thoracic diseases. Nevertheless, computer-aided diagnosis based on chest radiographs still faces considerable challenges in anomaly detection. Certain lesions in CXRs exhibit subtle radiographic characteristics with ambiguous boundaries, low pixel occupancy, and weak contrast. While existing studies primarily focus on improving multi-scale feature fusion, they frequently overlook complications arising from background noise and varied lesion morphology. This study introduces YOLOv11-MFF, an enhanced YOLOv11 network with three key innovations. Specifically, a novel Frequency-Adaptive Hybrid Gate (FAHG) is developed to improve contrast differentiation between lesions and background. A Multi Scale Parallel Large Convolution (MSPLC) block is designed and integrated with the original C3k2 module to expand receptive fields and enhance long-range modeling capacity. Furthermore, a Feature Fusion module (FF) is introduced to reinforce target-relevant feature representation through channel-wise modulation via weight recalibration mechanisms. Benefiting from these advancements, the network achieves significant improvements in detecting multi-scale and overlapping lesions. Experimental results on the public VinDr-CXR dataset demonstrate that YOLOv11-MFF outperforms state-of-the-art models, achieving a precision of 48.2%, recall of 42.5%, mAP@0.5 of 41.5%, and mAP@0.5:0.95 of 22.6%.
Bibliografia:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
content type line 23
ISSN:1932-6203
1932-6203
DOI:10.1371/journal.pone.0334283