Pixel-Wise Vessel Cross-Section Recognition at Bifurcations

The patent replaces conventional mechanical image processing algorithms with a deep learning-based semantic segmentation model. The learning model automatically identifies and segments different vascular structures (main trunk, side branch, bifurcated portion) and their lumen boundaries from medical images, achieving high recognition precision without complex manual processing steps.

The patent creates a digital copy of the medical image data and processes this copy through the learning model to generate segmented results. The model learns from training data containing labeled vascular structures and reproduces accurate segmentations by copying the patterns and features from the training images, enabling automatic discrimination of vascular components.

The patent performs preliminary training of the learning model using a comprehensive dataset of labeled medical images before actual use. During training, the model learns to identify vascular structures, their boundaries, and spatial relationships. This preliminary action enables the model to achieve high reliability in recognizing main trunk, side branch, and bifurcated portion structures when applied to new medical images.

The patent optimizes model parameters including learning rate, batch size, and network architecture to efficiently learn from training data. By adjusting these parameters, the model achieves high accuracy in vascular structure recognition while managing computational resource requirements and training data quantity effectively.