The three most pressing issues that are preventing #artificialintelligence from revolutionizing #ophthalmology, and how we aim to tackle it:
Problem 1: Data
Lack of real-world datasets due to proprietary and privacy concerns introduce difficulties when training data-hungry neural networks.
Problem 2: Interpretability
Inability to explain why an autonomous model has given an output or prediction. This “black box model” makes it difficult to build clinician and patient trust, even if a machine’s performance exceeds that of a human.
Problem 3: Computational Resources
Deep learning applications in areas such as image classification, segmentation, and object detection rely heavily on computing power and resources - creating an unsustainable model.
Using the novel distillation with no labels (DINO) algorithm to train vision transformers, pathology can be identified in real-time without requiring large domain-specific datasets.