Artificial intelligence in retinopathy of prematurity: transfer learning and federated learning

Artificial intelligence can help resolve the lack of specialists in retinopathy of prematurity (ROP) and its associated diagnostic subjectivity.Obstacles to the use of deep learning (DL) for ROP tasks include the condition's low prevalence, data paucity, and difficulty in multiinstitutional data sharing to optimize training.Transfer learning (TL) and federated learning (FL) are advanced strategies to address DL issues.This review highlights the advantages of TL and FL applications in various ROP-related tasks.TL and FL achieve outstanding results for ROP screening, triaging, and monitoring, with certain algorithms exceeding the baseline DL models.TL assists the construction of generalizable ROP models despite little data.FL lays the groundwork for safe data exchange between institutions in TL.However, both TL and FL entail shortcomings associated with inadequate generalizability and data privacy attacks.Further research is needed to address the unsolved interpretability and liability issues in TL and FL models.Although both TL and FL have great potential to overcome DL constraints and improve the diagnosis of ROP, more work is needed to address application concerns such as model interpretability and liability.