Image Encoder Training With Weighted Negatives for WSI Search
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing image encoder training methods using contrastive learning incorrectly assume negative samples in certain scenarios, leading to incorrect feature extraction and reduced precision in downstream search tasks for whole slide images.
Innovation Solution
The method involves cropping whole slide images into tissue images, generating feature vectors, clustering these vectors to identify key images, querying a database for target images, and assigning weights to negative samples to distinguish their negative degree, thereby improving the contrastive learning process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If contrastive learning is used to train the image encoder with standard positive and negative sample pairs, then the training process is simple and efficient, but the precision of extracted image features deteriorates due to wrong positive and negative sample selection
Solution Approach 1:
The patent applies local quality by differentiating the treatment of different negative samples based on their spatial relationship with the anchor image. Negative samples are divided into two categories: those from the same whole slide image (with lower weight) and those from different whole slide images (with higher weight). This localized differentiation of sample quality enables the model to learn from relevant negative samples while reducing the impact of irrelevant ones, thereby improving feature extraction precision without completely changing the training framework.
2Device complexity
If all negative samples are treated equally in contrastive learning, then the training process is straightforward, but the image encoder cannot accurately distinguish between different types of negative samples leading to wrong training direction
Solution Approach 1:
The patent implements local quality by assigning different weights to different negative samples based on their source. Negative samples from the same whole slide image as the anchor image are assigned a lower weight (first weight), while negative samples from different whole slide images are assigned a higher weight (second weight). This localized differentiation maintains the overall simplicity of contrastive learning while significantly improving training direction accuracy by preventing the model from being misled by irrelevant negative samples.
Solution Approach 2:
The patent applies parameter changes by modifying the loss function to include weight parameters that differentiate between types of negative samples. The weighted contrastive loss function introduces parameters (first weight and second weight) that change the contribution of different negative samples to the overall loss. This parameter modification enables the model to focus on more informative negative samples while reducing the impact of less informative ones, thereby improving training reliability without substantially increasing process complexity.
Data Source
AI summary
Provided is a method for searching for a whole slide image performed by a computer device, which relate to the field of artificial intelligence. The method includes: cropping a whole slide image into a plurality of tissue images; generating, through an image encoder, image feature vectors respectively corresponding to the plurality of tissue images; clustering the image feature vectors respectively corresponding to the plurality of tissue images, to determine at least one key image from the plurality of tissue images; querying, based on image feature vectors respectively corresponding to the at least one key image, a database to obtain at least one target image package corresponding to the at least one key image; and determining a whole slide image to which at least one candidate tissue image comprised in the at least one target image package respectively belongs as a final search result.