Automatic Retinoblastoma Screening and Surveillance Using Deep Learning

Ruiheng Zhang, Li Dong, Ruyue Li, Kai Zhang, Yitong Li, Hong-shu Zhao, Jitong Shi, Xin Ge, Xiaolin Xu, Libin Jiang, Xuhan Shi, Chuan Zhang, Wen‐Da Zhou, Liangyuan Xu, Haotian Wu, Heyan Li, Chuyao Yu, Jing Li, Jianmin Ma, Wenbin WeimedRxiv 2022 · DOI

Abstract

Retinoblastoma is the most common intraocular malignancy in childhood. With advanced management strategies, global salvage and overall survival have significantly improved, proposing subsequent challenges regarding long-term surveillance and offspring screening. Here, we developed a deep learning algorithm, called Deep Learning Assistant for Retinoblastoma (DLA-RB), trained on a total of 36,623 images from 713 patients. We validated it in a prospectively collected dataset comprised of 1,366 images from 139 eyes of 103 patients. In identifying active retinoblastoma tumors (treatment required) from all clinically suspected patients, the area under the receiver operating characteristic curve (AUC) of DLA-RB reached 0.991 (95% CI 0.970-1.000). In identifying active retinoblastoma from stable retinoblastoma patients (where treatment is not required), the AUC reached 0.962 (95% CI 0.915-1.000). Cost-utility analysis revealed that the DLA-RB-based diagnosis mode is more cost-effective in both retinoblastoma diagnosis and activity surveillance. The DLA-RB achieved high accuracy and sensitivity in identifying active retinoblastoma from normal and stable fundus images, suggesting it can be effectively incorporated into future telemedicine programs.

Introduction

Retinoblastoma (RB) represents the most prevalent intraocular malignancy in the pediatric population. While advancements in treatment have markedly increased survival rates, this success has introduced a secondary clinical challenge: the necessity for lifelong surveillance. Survivors and their offspring require frequent, high-quality ophthalmic examinations to detect potential relapses or new primary tumors.

Currently, the gold standard for monitoring involves "Examinations Under Anesthesia" (EUAs), conducted by specialized pediatric ophthalmologists. In many regions, particularly in developing nations, the scarcity of these specialists leads to delayed diagnoses and significant socioeconomic burdens on families who must travel to tertiary care centers. To address this, researchers developed DLA-RB (Deep Learning Assistant for Retinoblastoma), an automated system designed to distinguish between healthy eyes, stable (treated) lesions, and active tumors requiring intervention.

Methodology

The DLA-RB system utilizes a deep learning pipeline based on the ResNet-50 architecture. This model was selected following comparative testing against other architectures, such as InceptionResnetV3, due to its superior performance and computational efficiency.

The system was trained to execute two primary binary classification tasks:

  1. Screening: Differentiating active retinoblastoma from a normal fundus.
  2. Surveillance: Differentiating active (recurring or new) tumors from stable, treated lesions. Stable lesions often present with calcification or scarring that can mimic active disease, making this a high-difficulty task for automated systems.

To ensure clinical reliability, the researchers employed Grad-CAM (Gradient-weighted Class Activation Mapping). This technique generates heatmaps that visualize the specific regions of the retina the AI prioritizes, allowing clinicians to verify that the model is focusing on relevant pathology rather than image artifacts.

Dataset and Validation

The study leveraged an extensive dataset of fundus images:

  • Development Dataset: 36,623 images retrospectively collected from 713 patients at Beijing Tongren Hospital (2018–2022).
  • Prospective Validation Dataset: 1,366 images from 103 patients (139 eyes) collected in a real-world clinical setting in 2022.
  • Hardware Robustness: The prospective data included images from multiple generations of RetCam3 cameras to ensure the model's performance remained consistent across different hardware iterations.

Results and Clinical Impact

DLA-RB demonstrated performance that was non-inferior—and in several metrics, superior—to ophthalmologists with 2–5 years of experience.

Diagnostic Accuracy

In the identification of active tumors among clinically suspected patients, the model achieved an AUC of 0.991 (Sensitivity: 97.9%, Specificity: 100%). For the more complex task of activity surveillance (distinguishing active from stable lesions), the model achieved an AUC of 0.962. Notably, the AI surpassed competent human readers in sensitivity for the surveillance task, though it exhibited slightly lower specificity. In an oncological context, this bias toward flagging "false alarms" rather than missing active tumors is considered a safer clinical profile.

Cost-Utility Analysis

Beyond technical performance, the study conducted a formal cost-utility analysis using Markov models. The AI-integrated diagnostic mode was found to be significantly more cost-effective than the traditional expert-only model. By reducing unnecessary referrals and the need for frequent travel to specialized centers, the AI approach offers a sustainable path for scaling pediatric ophthalmic care.

Discussion and Limitations

While DLA-RB represents a significant step forward for telemedicine in pediatric oncology, several limitations remain. The study relied on data from a single center (Beijing Tongren Hospital), necessitating future multi-center validation to ensure global generalizability. Furthermore, the system's efficacy is dependent on the quality of the input images. Because RB often originates in the far periphery of the retina, the AI can only detect what is captured by the operator.

Future iterations of the software may include an "operation indicator" to guide users in capturing peripheral views and an automated quality-control module to reject blurred or low-quality images in real-time. Despite these challenges, DLA-RB provides a robust framework for automating the screening and monitoring of retinoblastoma, potentially transforming the standard of care for survivors worldwide.

Summary updated

Abstract

ABSTRACT Retinoblastoma is the most common intraocular malignancy in childhood. With the advanced management strategy, the global salvage and overall survival have significantly improved, which proposes subsequent challenges regarding long-term surveillance and offspring screening. Here, we developed deep learning algorithm, called Deep Learning Assistant for Retinoblastoma (DLA-RB) training on A total of 36623 images from 713 patients. We validated it in the prospectively collected dataset, comprised of 1366 images form 139 eyes of 103 patients. In identifying active retinoblastoma tumors (treatment required) from all clinical-suspected patients, the area under the receiver operating characteristic curve (AUC) of DLA-RB reached 0.991 (95% CI 0.970-1.000). In identifying active retinoblastoma from stable retinoblastoma patients (treatment is not required), AUC of DLA-RB reached 0.962 (95% CI 0.915-1.000), respectively. Cost-utility analysis revealed that DLA-RB based diagnosis mode is more cost-effective in both retinoblastoma diagnosis and retinoblastoma activity surveillance. The DLA-RB achieved high accuracy and sensitivity in identifying active retinoblastoma from the normal and stable retinoblastoma fundus. It can be incorporated within telemedicine programs in the future.
BibTeX
@misc{zhang2022automatic,
  author = {Ruiheng Zhang and Li Dong and Ruyue Li and Kai Zhang and Yitong Li and Hong-shu Zhao and Jitong Shi and Xin Ge and Xiaolin Xu and Libin Jiang and Xuhan Shi and Chuan Zhang and Wen‐Da Zhou and Liangyuan Xu and Haotian Wu and Heyan Li and Chuyao Yu and Jing Li and Jianmin Ma and Wenbin Wei},
  title = {{Automatic Retinoblastoma Screening and Surveillance Using Deep Learning}},
  year = {2022},
  doi = {10.1101/2022.08.23.22279103},
  howpublished = {medRxiv},
}
Figure