Abstract
Infantile cataract is a leading cause of childhood blindness. While AI diagnostic systems show promise, their generalizability is often limited by the heterogeneity of multicenter slit-lamp images. This study developed two lens partition strategies (LPSs) based on Faster R-CNN and Hough transform to improve detection generalizability. Using 1,643 multicenter images, the Faster R-CNN LPS achieved an average intersection over union (IoU) of 0.9419 and 0.9107 for normal and abnormal lenses, respectively. Compared to the Hough transform, Faster R-CNN improved grading accuracy by 5.31% and reduced processing time from 34.46s to 0.25s per image. This system demonstrates significant potential for real-world clinical application across diverse ophthalmic centers.
Introduction
Infantile cataracts represent a critical global health challenge. Although deep learning models have achieved high performance in controlled, single-center environments, they frequently struggle when deployed across multiple clinics. This lack of generalizability is primarily due to "noise" in slit-lamp images—such as eyelashes, eyelids, and varying lighting conditions—which differs significantly between clinical sites. To overcome this, we propose a standardized Lens Partition Strategy (LPS) to accurately isolate the lens from surrounding ocular noise before diagnostic classification.
Methods
The researchers developed a two-stage framework consisting of an automated lens partition stage followed by a screening and grading stage.
Study Design and Datasets
The study utilized 1,643 slit-lamp images: 886 from the CC-Cruiser dataset (single center) and 757 from a multicenter dataset across four independent clinics. The experimental workflow involved evaluating different LPS methods and determining the minimal multicenter training data required for stable performance.
Technical Architecture
Two LPS methods were compared: the traditional geometric Hough Transform and a deep learning-based Faster R-CNN. The Faster R-CNN model was trained to detect the lens as an object, providing greater robustness against non-circular lens shapes and lighting artifacts. Following partition, a 50-layer ResNet (ResNet-50) was used to perform screening (normal vs. cataract) and three-degree grading (area, density, and location).
Results
The Faster R-CNN-based LPS significantly outperformed the Hough Transform in both accuracy and efficiency.
- Accuracy: Faster R-CNN achieved a mean IoU of 0.9419 (normal) and 0.9107 (abnormal), whereas the Hough Transform achieved 0.8970 and 0.8138.
- Efficiency: Faster R-CNN processed images in 0.25 seconds, a 137x speedup over the 34.46 seconds required by the Hough Transform.
- Diagnostic Performance: The integrated pipeline achieved an AUC of 0.997 for screening and high AUCs (0.965–0.979) for all grading tasks.
- Data Requirements: Performance stabilized after adding just 2-3 subsets of multicenter data, indicating that capturing equipment variance is more important than total sample volume.
Discussion
This study demonstrates that the primary bottleneck for AI generalizability in infantile cataract detection is the preprocessing stage. By implementing a robust, deep learning-based lens partition strategy, we can effectively filter out clinical noise and standardize inputs for classification models. While the system is highly effective for whole-lens isolation, future iterations will focus on pixel-level lesion segmentation and the integration of multimodal imaging to capture posterior lens pathologies. This approach paves the way for the deployment of reliable AI diagnostic tools in diverse real-world clinical settings.
Summary updated