Figure 2.

Examples of heterogeneous training splits of the diabetic retinopathy dataset with 4 participating institutions consisting of 20 total images (10 healthy and 10 diseased). (A) Training split with sample size variability; institutions 1, 2, 3, and 4 have 2, 4, 6, and 8 samples respectively, and the data at each institution have a balanced label distribution. (B) Training split with label distribution variability; each institution has equal number of data samples but varying label distribution.