TL;DR
This paper introduces ACM-Assessor, an AI system that uses ultrasound biomicroscopy (UBM) images to automatically assess the mechanisms behind primary angle-closure glaucoma (PACG). The system achieves expert-level accuracy, significantly aids novice practitioners, and operates much faster than human experts, paving the way for more efficient and accessible PACG diagnosis and treatment.
What Problem Does This Paper Address?
Primary angle-closure glaucoma (PACG) is a leading cause of blindness, particularly in Asia. Its progression is insidious and irreversible, making early and accurate diagnosis crucial. The key to effective treatment lies in precisely identifying the underlying mechanisms of angle closure (AC), such as pupillary block (PB), thick peripheral iris (TPI), or an anteriorly located ciliary body (ALCB). Often, multiple mechanisms (MM) are present. While gonioscopy is a standard diagnostic tool, it has limitations. Ultrasound biomicroscopy (UBM) offers superior visualization of posterior structures but its interpretation is subjective, time-consuming, and requires significant expertise. Existing AI approaches for AC mechanism assessment often rely on different imaging modalities (like AS-OCT), have limited validation, or cannot fully assess all relevant structures, leaving a gap for UBM-based AI solutions.
Methods
The ACM-Assessor system is built upon several key components:
- Image Preprocessing: This involves converting UBM image pixel values to numerical representations and using a UNet++ network for segmenting critical anterior chamber structures (iris, sclera, ciliary body).
- Scleral Spur Localization: A Python program automatically identifies the scleral spur, a crucial anatomical landmark.
- Binary Classification Models: Three separate models, leveraging Support Vector Machines (SVM) and Random Forests (RF), are trained to detect the presence of individual AC mechanisms:
- Pupillary Block (PB): Based on iris root insertion and curvature.
- Thick Peripheral Iris (TPI): Measured by distances between iris intersections with specific circles around the scleral spur.
- Anteriorly Located Ciliary Body (ALCB): Assessed by the contact length between the anterior ciliary body and the posterior iris.
- Integrated Assessment Model: This model combines the outputs of the binary classifiers to categorize AC mechanisms into four types: pure PB, pure non-PB, multiple mechanisms (MM), and others. This is achieved using SVM and RF algorithms.
The system's development involved extensive feature extraction and custom classifier design, aiming to capture specific UBM image characteristics indicative of each mechanism.
Datasets
The study utilized a substantial dataset of 8482 UBM images collected retrospectively from 1160 patients across two institutions: Renmin Hospital of Wuhan University (RHWU) and Huangshi Aier Eye Hospital (HAEH).
- Training and Internal Testing: A portion of the RHWU dataset was used for training and internal validation. Specific subsets were allocated for each binary classification task (PB, TPI, ALCB).
- External Testing: 2266 UBM images from HAEH served as an external testing set to evaluate the model's generalizability.
- Independent Testing Set: An additional 436 images from 436 eyes of 299 patients were used for a rigorous independent evaluation, including comparisons with human experts and assessment of its utility for training beginners.
Key Insights and Ideas
- UBM as a Rich Source for AI: The study highlights UBM's unique capability to visualize posterior structures, making it ideal for comprehensive AC mechanism assessment, a capability often limited in other imaging modalities.
- Beyond Single Mechanism Detection: Unlike many prior AI systems that focus on identifying only one mechanism per image, ACM-Assessor's ability to detect multiple mechanisms (MM) is crucial, as MM is common in PACG and impacts treatment.
- Automating Subjective Tasks: By automating scleral spur localization and AC mechanism classification, the system aims to overcome the subjectivity and variability inherent in manual UBM image interpretation.
- Bridging the Expertise Gap: The system not only performs at an expert level but also significantly improves the accuracy and efficiency of novice practitioners, suggesting its potential as a powerful educational tool.
Key Results
- High Accuracy in Binary Classification: ACM-Assessor achieved impressive accuracies: 0.924 for PB, 0.925 for TPI, and 0.947 for ALCB on the internal testing set. External testing also showed strong performance (e.g., 0.883 for PB).
- Robust Integrated Assessment: The integrated assessment model achieved an accuracy of 0.839 on the internal test set.
- Expert-Level Performance: In direct comparison with three senior ophthalmologists, ACM-Assessor's accuracy was comparable (p > 0.05) across all assessed categories.
- Significant Speed Advantage: The system operated approximately four times faster than human experts, completing evaluations in 784.80 seconds compared to the experts' average of 3020.39 seconds.
- Enhanced Beginner Performance: With model assistance, beginners' accuracy improved by 0.117 for binary classification and 0.219 for integrated assessment, demonstrating its value as a training aid.
Limitations and Future Work
- Single Image Analysis: The current model analyzes AC mechanisms from a single UBM image. Future work could involve analyzing multiple images from the same patient for a more comprehensive assessment.
- Device Specificity: The model was trained on UBM images from a single device. Transfer learning could be explored to adapt the model for images from different UBM brands.
- Limited Mechanism Scope: The study focused on PB, TPI, and ALCB. Other AC mechanisms, such as plateau iris, were not explicitly identified, though plateau iris is related to ALCB. Future research could expand the system to include these.
Summary updated