Purpose: To conduct a demonstration experiment of a multi-facility artificial intelligence (AI) model built using federated learning that does not bring medical image out of the facility.
Subjects and methods: We used federated learning to create an AI model that automatically extracts recess in the optic nerve head and its limbus. Graphics processing unit (GPU) -equipped machines were placed in three facilities (Tsukazaki Hospital, Yamaike Eye Clinic, and Niimi Eye Institute) and connected to the central server via a dedicated line. A total of 356 fundus photographs were collected in each facility, and optometrists at each facility used image processing software to distinguish the depressed portion of the optic nerve papilla from the peripheral part by color under the guidance of a glaucoma specialist. The AI model consisted of three steps: the extraction of the entire area of the optic nerve papilla (model P), identification of the recessed area (model C), and identification of the marginal area (model D). The AI model learned at the third period, imitating the environment where images are accumulated in order. Images were rotated 5 to 20 times until the loss values converged in each period. Only the parameters obtained were sent to the central server at each rotation to create a weight ed average model. The weighted average model during the rotation with the lowest loss value was returned to each facility as the integrated model for each period and used as the initial value at the next period.
Results: The loss values at each period in each model were as follows: Model P (79.57 at the 1st period, 61.12 at the 2nd period, and 57.90 at the 3rd period), Model C (0.423 at the 1st period, 0.397 at the 2nd period, and 0.364 at the 3rd period), and Model D (0.259 at the 1st period, 0.212 at the 2nd period, and 0.183 at the 3rd period).
Conclusion: The loss value improved steadily in the federated learning process.
Nippon Ganka Gakkai Zasshi (J Jpn Ophthalmol Soc) 126: 421-435,2022.