Whole Slide Image Encoder Training With Weighted Negative Samples

The patent applies preliminary action by performing clustering analysis on negative samples before the contrastive learning training process. The method pre-calculates similarity metrics and identifies high-similarity negative samples that should be treated as positive samples, then adjusts the training data accordingly. This preliminary preparation ensures that the image encoder receives correctly labeled training data from the start, preventing the propagation of errors throughout the training process while maintaining relatively simple implementation.

The patent applies local quality by differentiating the treatment of negative samples based on their individual characteristics. Instead of uniformly processing all negative samples, the method calculates similarity metrics for each negative sample relative to the anchor image, identifies high-similarity samples through clustering, and applies different weights or treatments accordingly. This localized differentiation ensures that high-similarity negative samples (which should be positive samples) receive appropriate attention, improving feature extraction precision while adding manageable complexity to the training process.