FusionNet OCT Ensemble for Alzheimer's Disease Classification

The patent divides the deep learning model into multiple independent baseline models (e.g., ResNet50, VGG16, InceptionV3) that process different features separately. Each baseline model is trained independently on specific feature sets (RNFL thickness, GCIPL thickness, deviation maps), and their predictions are aggregated through ensemble voting. This segmentation prevents overfitting by avoiding a single monolithic complex model while maintaining high detection accuracy through combined insights from multiple specialized models.

The patent merges multiple baseline models into an ensemble system where predictions from different models are combined through voting mechanisms. The FusionNet architecture integrates features from multiple sources (thickness maps, deviation maps, en face images) and combines predictions from multiple baseline models to produce a final diagnosis. This merging approach leverages the strengths of each individual model while reducing the overfitting risk associated with any single complex model.

The patent makes the OCT device multi-functional by training a single unified model to process multiple types of OCT data (RNFL thickness maps, GCIPL thickness maps, deviation maps, and en face images) for AD detection. This universal approach allows one OCT device to perform multiple diagnostic functions without requiring additional invasive imaging modalities, thereby maintaining high detection accuracy while reducing system complexity and cost.

The patent creates digital copies of retinal structures through OCT imaging and uses these digital representations for diagnosis rather than requiring multiple physical invasive procedures. The deep learning model processes these digital copies (images and derived features) to detect AD, replacing the need for multiple invasive imaging modalities with a single non-invasive OCT scan that provides sufficient diagnostic information through computational analysis.

The patent segments the model complexity by processing different OCT inputs through separate baseline models before aggregation. Each input type (thickness maps, deviation maps, en face images) is processed by dedicated baseline models that are independently trained and optimized. This segmentation allows the system to handle multiple inputs without creating a single monolithic complex model, as each component remains relatively simple and modular.

The patent introduces an intermediary aggregation layer that combines predictions from multiple baseline models processing different OCT inputs. This intermediary voting mechanism simplifies the overall system architecture by providing a clear, straightforward method for integrating multiple sources of information. The intermediary layer acts as a mediator that reconciles different model predictions through voting, avoiding the need for complex integration mechanisms while maintaining high predictive performance.