Inference Batch Normalization for MIL CNN Slide Labeling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional machine-learning models for predicting slide-level class labels in whole-slide histopathology images face performance issues due to their unsuitability for handling bags of patches from the same image, leading to diminished prediction accuracy.
Innovation Solution
Generating inference-phase-specific batch normalization parameters for a multiple-instance learning convolutional neural network (MILCNN) to normalize features during the inference phase, rather than relying on training-phase-determined running means and variances, allowing the model to better predict slide-level class labels like gene alterations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine-learning models use running mean and variance from training phase for batch normalization during inference, then the model structure remains simple, but the prediction accuracy diminishes because the parameters are not fitted to the specific whole-slide image data
Solution Approach 1:
The patent pre-calculates batch normalization parameters (mean and variance) specifically for the inference phase by processing all patches from the whole-slide image before making predictions. This preliminary calculation ensures the normalization parameters are fitted to the specific image data, improving prediction accuracy without adding complexity during the actual inference process.
Solution Approach 2:
The patent makes the batch normalization parameters dynamic by calculating them specifically for each inference case rather than using fixed training-phase parameters. The parameters adapt to the specific whole-slide image being analyzed, allowing the model to optimize its normalization for each prediction task while maintaining a relatively simple overall structure.
2Reliability
If the model processes bags of patches from the same whole-slide image, then it can capture slide-level biomarkers, but traditional batch normalization techniques become unsuitable and reduce prediction performance
Solution Approach 1:
The patent applies local quality by calculating batch normalization parameters specifically for the inference phase data rather than using global training-phase parameters. Each whole-slide image receives customized normalization parameters based on its own patch statistics, making the normalization process adapted to the specific MIL data structure and improving prediction reliability.
3Measurement precision
If inference-phase-specific batch normalization parameters are generated, then the model achieves better prediction accuracy for slide-level class labels, but additional computation is required during the inference phase
Solution Approach 1:
The patent performs the computationally intensive batch normalization parameter calculation as a preliminary step before actual prediction. By pre-processing the patches and calculating means and variances beforehand, the model achieves high prediction accuracy while keeping the actual inference time relatively short, as the heavy computation is done in advance.
Data Source
AI summary
A method implemented by one or more processors includes segmenting an image into a plurality of patches grouping the plurality of patches into at least one bag of patches, and inputting the at least one bag of patches into a machine-learning model trained to generate a prediction of an image class label based on the at least one bag of patches. The machine-learning model includes a first layer trained to generate one or more feature maps based on the at least one bag of patches, a second layer trained to normalize the one or more feature maps utilizing a set of batch normalization parameters determined from the at least one bag of patches to generate one or more normalized feature maps, and a third layer trained to generate the prediction of the image class label based at least in part on the one or more normalized feature maps.