A deep learning based smartphone application for early detection of nasopharyngeal carcinoma using endoscopic images
Deep Learning Medical ImagingPosted by mhb on 2025-11-06 18:26:27 |
Share: Facebook | Twitter | Whatsapp | Linkedin Visits: 225
The study introduces “Nose-Keeper”, a smartphone-based deep learning application designed for the early detection of nasopharyngeal carcinoma (NPC) and other nasal diseases using endoscopic images.
Methods
-
Dataset: 39,340 nasal endoscopic white-light images collected from 3 NPC high-incidence centers.
-
Models Tested: Eight advanced deep learning architectures (e.g., Swin Transformer, ResNet, DenseNet, ConvNeXt).
-
Best Model: Swin Transformer (SwinT)—selected for deployment due to highest accuracy and robustness.
-
Evaluation: Compared performance with nine experienced otolaryngologists and on external datasets.
Key Results
-
Overall Accuracy: 92.27% (95% CI: 90.66–93.61%)
-
NPC Detection Sensitivity: 96.39%
-
NPC Specificity: 99.91%
-
Outperformed all human experts in NPC diagnosis.
-
The model remained stable under variations in brightness, rotation, and blur.
-
Integrated Grad-CAM explainable AI to visualize lesion areas for safer decision-making.
App Functionality
-
Connects to a nasal endoscope or imports stored images.
-
Performs AI-based diagnosis in 0.5–1.1 seconds per image.
-
Displays diagnostic results and heatmaps for transparency.
-
Provides educational content and reference images for users and clinicians.
Discussion & Significance
-
First-ever smartphone AI app for NPC detection.
-
Offers low-cost, rapid, and accessible screening, especially valuable for low- and middle-income regions with few specialists.
-
Can reduce diagnostic delays, assist primary healthcare providers, and raise public awareness of nasal health.
Limitations
-
Lack of prospective clinical validation.
-
Dependent on internet access and image quality.
-
Data collected using professional devices—performance on household endoscopes needs testing.
Conclusion
Nose-Keeper demonstrates that AI-powered mobile tools can accurately and efficiently detect nasopharyngeal carcinoma, potentially transforming early cancer screening in resource-limited settings. Future improvements include on-device inference, image quality control, and broader disease recognition.
Search
Recent News
- Clinical application research on the titanium metal metatarsal prosthesis designed through FEA and manufactured by 3D printing
- Manually weighted taxonomy classifiers improve species-specific rumen microbiome analysis compared to unweighted or average weighted taxonomy classifiers
- A deep learning based smartphone application for early detection of nasopharyngeal carcinoma using endoscopic images
- sCIN: a contrastive learning framework for single-cell multi-omics data integration
- CustOmics: A versatile deep-learning based strategy for multi-omics integration
- Tracking temporal progression of benign bone tumors through X-ray based detection and segmentation
- mCNN-GenEfflux: enhanced predicting Efflux protein and their super families by using generative proteins combined with multiple windows convolution neural networks
- Evaluation of normalized T1 signal intensity obtained using an automated segmentation model in lower leg MRI as a potential imaging biomarker in Charcot– Marie–Tooth disease type 1A
Popular Paper
- sCIN: a contrastive learning framework for single-cell multi-omics data integration
- A deep learning based smartphone application for early detection of nasopharyngeal carcinoma using endoscopic images
- CustOmics: A versatile deep-learning based strategy for multi-omics integration
- Clinical application research on the titanium metal metatarsal prosthesis designed through FEA and manufactured by 3D printing
- Manually weighted taxonomy classifiers improve species-specific rumen microbiome analysis compared to unweighted or average weighted taxonomy classifiers