A mobile app for Glaucoma diagnosis and its possible clinical applications

Fan Guo, Weiqing Li, Xin Zhao, Junfeng Qiu, Yuxiang MaiBMC Medical Informatics Decis. Mak 2020 · DOI

TL;DR

This paper introduces "Yanbao," a mobile application designed for convenient glaucoma diagnosis. Leveraging smartphone fundus cameras and a novel U-Net-based deep learning model, Yanbao achieves promising accuracy in detecting glaucoma, aiming to improve accessibility and efficiency in healthcare.

What Problem Does This Paper Address?

Glaucoma is a leading cause of irreversible blindness, and early detection and treatment are crucial to mitigating its impact. However, access to specialized medical care, particularly in remote areas or regions with limited resources, can be a significant barrier. Traditional diagnostic methods often require specialized equipment and trained professionals, making them inaccessible to many. This paper addresses the need for a more accessible, user-friendly, and efficient tool for glaucoma screening that can be utilized by a wider population, potentially bridging the gap in healthcare access and alleviating the burden on medical professionals.

Methods

The core of the Yanbao app's diagnostic capability lies in its novel deep learning architecture and image processing techniques.

  • U-Net + CP + FL Architecture: The authors developed a modified U-Net architecture, enhanced with a "concatenating path" (CP) and a "fusion loss" (FL) function. The CP allows for better feature sharing and multi-scale input integration, while the FL combines weighted binary cross-entropy and dice coefficient loss to address data imbalance issues common in medical imaging. This architecture is designed for robust segmentation of the optic disc (OD) and optic cup (OC) from retinal images.
  • OD Localization: A multi-step process is employed to accurately locate the optic disc. This involves image enhancement, brightest region extraction, and blood vessel segmentation. A sliding window approach with confidence calculation is then used to pinpoint the OD.
  • CDR and NRR Measurement: Once the OD and OC are segmented, key clinical parameters like the cup-to-disc ratio (CDR) and neuro-retinal rim (NRR) are calculated. These parameters are crucial for glaucoma screening.
  • Feature Selection and Classification: To improve classification accuracy, 25 features derived from OD, OC, and NRR analysis are extracted. A feature selection process involving variance analysis, correlation analysis, and random forest ranking identifies the eight most significant features. A Gradient Boosting Decision Tree (GDBT) classifier, combined with SMOTE (Synthetic Minority Over-sampling Technique) for handling imbalanced data, is used for the final glaucoma classification.
  • Mobile Application: The Yanbao app itself is designed for ease of use, allowing users to capture retinal images using smartphone fundus cameras (like the oDocs Fundus) or upload existing ones. It provides a user interface for image acquisition, data upload, and displaying diagnostic results.

Datasets

  • ORIGA Dataset: This public retinal image database was used for training and evaluating the segmentation performance of the U-Net + CP + FL model.
  • DRISHTI-GS1 Dataset: This dataset was used to evaluate the prediction accuracy of the GDBT+SMOTE classifier.
  • Clinical Setting Data: The authors collected data from 274 patients (648 fundus images) from the Second Xiangya Hospital of Central South University for real-world clinical assessment of the app's performance.

Key Insights and Ideas

  • Bridging the Healthcare Gap: The core idea is to democratize glaucoma screening by making it accessible via a smartphone app, reducing reliance on specialized clinics and professionals.
  • Novel Deep Learning Architecture: The U-Net + CP + FL architecture represents a significant advancement in adapting deep learning for medical image segmentation, particularly for datasets with limited samples.
  • Importance of Data Imbalance Handling: The study strongly emphasizes the critical role of techniques like SMOTE and class weighting in achieving robust glaucoma classification, especially given the inherent imbalance between healthy and diseased cases.
  • Comprehensive Clinical Parameter Extraction: The app doesn't just provide a binary diagnosis but extracts clinically relevant parameters (CDR, NRR) that can aid clinicians in their assessment.
  • User-Centric Design: The app's interface is designed to be intuitive and user-friendly, considering potential users who may have limited technical or medical knowledge.

Key Results

  • Segmentation Performance: The proposed U-Net + CP + FL model achieved improved segmentation performance (Sdisc, Accdisc, Scup, Acccup, SCDR) compared to the standard U-Net on the ORIGA dataset.
  • Classification Performance: The GDBT+SMOTE classifier demonstrated strong performance on the DRISHTI-GS1 dataset, achieving high accuracy (0.806) and AUC (0.878).
  • Clinical Test Results: In a clinical setting, the Yanbao app achieved an overall accuracy of 0.7731, with a sensitivity of 0.7860 and specificity of 0.7654 on 274 patients. This performance is comparable to results obtained on public datasets.
  • User Experience: Usability tests indicated a high average SUS score of 80.8, suggesting good applicability and a rational user interface design.
  • Response Time: The average response time for the app was 9.18 seconds, which is considered acceptable given the cloud-based processing.

Limitations and Future Work

  • Malignant Glaucoma: The study notes that the app's detection of subclasses of secondary glaucoma does not include malignant glaucoma.
  • Data Imbalance: While SMOTE and class weighting were employed, the inherent data imbalance in medical datasets remains a challenge.
  • Clinical Interpretation: The app is positioned as an aid to doctors, not a replacement. Doctor's diagnosis is still required for definitive confirmation.
  • Future Work: The authors suggest that the app can be further developed to assist in assessing eye conditions beyond glaucoma, indicating a broader potential for mobile health solutions. The study also highlights the need for continued research into improving the accuracy and scope of mobile diagnostic tools.

Summary updated

BibTeX
@article{guo2020mobile,
  author = {Fan Guo and Weiqing Li and Xin Zhao and Junfeng Qiu and Yuxiang Mai},
  title = {{A mobile app for Glaucoma diagnosis and its possible clinical applications}},
  journal = {BMC Medical Informatics Decis. Mak.},
  year = {2020},
  doi = {10.1186/s12911-020-1123-2},
}
Figure