The present study aimed to conduct a real-time automatic analysis of two important surgical phases, which are continuous curvilinear capsulorrhexis (CCC), nuclear extraction, and three other surgical phases of cataract surgery using artificial intelligence technology. A total of 303 cases of cataract surgery registered in the clinical database of the Ophthalmology Department of Tsukazaki Hospital were used as a dataset. Surgical videos were downsampled to a resolution of 299 × 168 at 1 FPS to image each frame. Next, based on the start and end times of each surgical phase recorded by an ophthalmologist, the obtained images were labeled correctly. Using the data, a neural network model, known as InceptionV3, was developed to identify the given surgical phase for each image. Then, the obtained images were processed in chronological order using the neural network model, where the moving average of the output result of five consecutive images was derived. The class with the maximum output value was defined as the surgical phase. For each surgical phase, the time at which a phase was first identified was defined as the start time, and the time at which a phase was last identified was defined as the end time. The performance was evaluated by finding the mean absolute error between the start and end times of each important phase recorded by the ophthalmologist as well as the start and end times determined by the model. The correct response rate of the cataract surgical phase classification was 90.7% for CCC, 94.5% for nuclear extraction, and 97.9% for other phases, with a mean correct response rate of 96.5%. The errors between each phase’s start and end times recorded by the ophthalmologist and those determined by the neural network model were as follows: CCC’s start and end times, 3.34 seconds and 4.43 seconds, respectively and nuclear extraction’s start and end times, 7.21 seconds and 6.04 seconds, respectively, with a mean of 5.25 seconds. The neural network model used in this study was able to perform the classification of the surgical phase by only referring to the last 5 seconds of video images. Therefore, our method has performed like a real-time classification.

We aimed to predict keratoconus progression and the need for corneal crosslinking (CXL) using deep learning (DL). Two hundred and seventy-four corneal tomography images taken by Pentacam HR^® (Oculus, Wetzlar, Germany) of 158 keratoconus patients were examined. All patients were examined two times or more, and divided into two groups; the progression group and the non-progression group. An axial map of the frontal corneal plane, a pachymetry map, and a combination of these two maps at the initial examination were assessed according to the patients’ age. Training with a convolutional neural network on these learning data objects was conducted. Ninety eyes showed progression and 184 eyes showed no progression. The axial map, the pachymetry map, and their combination combined with patients’ age showed mean AUC values of 0.783, 0.784, and 0.814 (95% confidence interval (0.721–0.845) (0.722–0.846), and (0.755–0.872), respectively), with sensitivities of 87.8%, 77.8%, and 77.8% ((79.2–93.7), (67.8–85.9), and (67.8–85.9)) and specificities of 59.8%, 65.8%, and 69.6% ((52.3–66.9), (58.4–72.6), and (62.4–76.1)), respectively. Using the proposed DL neural network model, keratoconus progression can be predicted on corneal tomography maps combined with patients’ age.

Background Artificial intelligence (AI) in medical imaging diagnostics has huge potential, but human judgement is still indispensable. We propose an AI-aided teaching method that leverages generative AI to train students on many images while preserving patient privacy. Methods A web-based course was designed using 600 synthetic ultra-widefield (UWF) retinal images to teach students to detect disease in these images. The images were generated by stable diffusion, a large generative foundation model, which we fine-tuned with 6285 real UWF images from six categories: five retinal diseases (age-related macular degeneration, glaucoma, diabetic retinopathy, retinal detachment and retinal vein occlusion) and normal. 161 trainee orthoptists took the course. They were evaluated with two tests: one consisting of UWF images and another of standard field (SF) images, which the students had not encountered in the course. Both tests contained 120 real patient images, 20 per category. The students took both tests once before and after training, with a cool-off period in between. Results On average, students completed the course in 53 min, significantly improving their diagnostic accuracy. For UWF images, student accuracy increased from 43.6% to 74.1% (p<0.0001 by paired t-test), nearly matching the previously published state-of-the-art AI model’s accuracy of 73.3%. For SF images, student accuracy rose from 42.7% to 68.7% (p<0.0001), surpassing the state-of-the-art AI model’s 40%. Conclusion Synthetic images can be used effectively in medical education. We also found that humans are more robust to novel situations than AI models, thus showcasing human judgement’s essential role in medical diagnosis.

The present study aims to describe the use of machine learning (ML) in predicting the occurrence of postoperative refraction after cataract surgery and compares the accuracy of this method to conventional intraocular lens (IOL) power calculation formulas. In total, 3331 eyes from 2010 patients were assessed. The objects were divided into training data and test data. The constants for the IOL power calculation formulas and model training for ML were optimized using training data. Then, the occurrence of postoperative refraction was predicted using conventional formulas, or ML models were calculated using the test data. We evaluated the SRK/T formula, Haigis formula, Holladay 1 formula, Hoffer Q formula, and Barrett Universal II formula (BU-II); similar to ML methods, we assessed support vector regression (SVR), random forest regression (RFR), gradient boosting regression (GBR), and neural network (NN). Among the conventional formulas, BU-II had the lowest mean and median absolute error of prediction. Therefore, we compared the accuracy of our method with that of BU-II. The absolute errors of some ML methods were lower than those of BU-II. However, no statistically significant difference was observed. Thus, the accuracy of our method was not inferior to that of BU-II.