The study evaluates the classification performance of deep learning models for six types of fundus diseases in OCT images using a multi-source training strategy. They created a combined dataset of 6,165 images including age-related macular degeneration (AMD), diabetic macular edema (DME), retinal artery occlusion (RAO), retinal vein occlusion (RVO), epiretinal membrane (ERM), vitreomacular interface (VID), and normal controls (NO). They compared two strategies: training only on public OCTDL data (S1) and joint training with local clinical data from Shanxi Eye Hospital (S2), with testing on the OCTDL test set. Strategy S1 had limited performance due to small samples of some categories. The S2 strategy significantly improved the overall performance of all six deep learning models tested. The best was the ViT-Base model in S2 with an accuracy of 93.61%, a misdiagnosis rate of RAO of 0%, AMD of 1.34%, and RVO reduced from 14.89% to 8.51%. In particular, this strategy strengthens the recognition of certain diseases and affects the transfer effect of learning through model selection.