研究目的
To assess the performance of a deep-learning system to detect papilledema from fundus images taken at many international centers, from patients with a variety of ethnic backgrounds, types of fundus pigmentation, and ages and using a variety of commercially available digital fundus cameras.
研究成果
An artificial-intelligence, deep-learning algorithm trained on ocular fundus photographs had high sensitivity and specificity for discriminating between papilledema and normal optic nerves. Negative predictive values were high, but positive predictive values varied depending on the prevalence of papilledema in the population being studied. Further investigation is required to prospectively validate the use of deep-learning systems in various settings.
研究不足
The study was retrospective, resulting in an imbalance in class distribution among groups, a mix of consecutive series of patients and convenience samples, and labeling errors. The abnormal photographs were obtained after pharmacologic dilation of the pupils and may not reflect general practice. The network was trained and calibrated primarily to identify normal optic nerves and those with papilledema, with a low threshold for diagnosing papilledema to avoid false negatives.
1：Experimental Design and Method Selection
The study involved training, validating, and externally testing a deep-learning system to classify optic disks as being normal or having papilledema or other abnormalities from retrospectively collected ocular fundus photographs.
2：Sample Selection and Data Sources
A total of 15,846 fundus photographs from 7532 patients were used, with 14,341 photographs from 19 sites in 11 countries for training and validation, and 1505 photographs from 5 other sites for external testing.
3：List of Experimental Equipment and Materials
Various commercial digital fundus cameras were used for image acquisition after pharmacologic pupillary dilation.
4：Experimental Procedures and Operational Workflow
Images were centered on either the macula or the optic disk, always including the optic disk, at various fields of view. The deep-learning system consisted of a segmentation network (U-Net) to detect the location of the optic disk and a classification network (DenseNet) to classify the optic disk.
5：Data Analysis Methods
Performance was evaluated by calculating the area under the receiver-operating-characteristic curve (AUC), sensitivity, and specificity, compared with a reference standard of clinical diagnoses by neuro-ophthalmologists.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
