Neural Network Training with Mixed Quality Medical Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Supervised training of neural networks for medical image analysis faces challenges due to insufficient high-quality labeled training samples, leading to poor prediction accuracy, especially when mixed label quality is used, and the risk of overfitting.
Innovation Solution
A computer-implemented method for training artificial neural networks that involves acquiring two sets of training samples with differing label quality, where the neural network is trained using a cost function with a second part that prevents overfitting by setting an upper bound for the average prediction performance on a subset of high-quality samples, ensuring accurate predictions for critical medical images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If training samples with mixed label quality are used to increase the number of training samples, then the quantity of training data is improved, but the prediction accuracy deteriorates due to inaccurately labelled samples
Solution Approach 1:
The training samples are segmented into two distinct sets: a first set with high-quality labels and a second set with lower-quality labels. This segmentation allows the neural network to be trained on a larger quantity of samples while maintaining accuracy by treating different quality samples differently through separate cost function components.
Solution Approach 2:
Different quality levels are assigned to different portions of the training data. The cost function applies different weighting or treatment to samples from the first set versus the second set, ensuring that high-quality samples have a stronger influence on the training process while still incorporating lower-quality samples to increase overall data quantity.
2Productivity
If training is performed on all training samples without selection, then the productivity of training is improved, but the prediction accuracy deteriorates due to overfitting on low-quality labels
Solution Approach 1:
The training process is segmented into two parallel training operations: one on the first set of high-quality samples and another on the second set of lower-quality samples. This allows efficient use of all available data while preventing overfitting through separate optimization processes with differentiated cost function components.
Solution Approach 2:
The method applies partial action by not treating all samples equally - instead, it applies different training intensities or weightings to different sample sets. The cost function includes separate parts that apply different levels of influence to high-quality versus lower-quality samples, preventing excessive influence from inaccurate labels.
3Measurement precision
If a balanced drawing of training samples from two label-quality categories is used, then the prediction accuracy is improved, but the control over performance on high-quality samples is lost
Solution Approach 1:
The cost function incorporates feedback mechanisms through its two distinct parts: one part monitors performance on high-quality samples and another part monitors performance on lower-quality samples. This structured feedback allows explicit control over performance on high-quality samples while maintaining overall prediction accuracy through the combined optimization process.
Solution Approach 2:
The cost function parameters are changed to reflect different quality levels - separate cost function parts with different weights or optimization targets are applied to different sample sets. This parameter differentiation enables explicit control over performance on high-quality samples while still utilizing lower-quality samples for overall accuracy improvement.
Data Source
AI summary
The invention relates to a method (100) for supervised training of an artificial neural network for medical image analysis. The method comprises acquiring (SI) first and second sets of training samples, wherein the training samples comprise feature vectors and associated predetermined labels, the feature vectors being indicative of medical images and the labels pertaining to anatomy detection, to semantic segmentation of medical images, to classification of medical images, to computer-aided diagnosis, to detection and/or localization of biomarkers or to quality assessment of medical images. The accuracy of predetermined labels may be better for the second set of training samples than for the first set of training samples. The neural network is trained (S3) by reducing a cost function, which comprises a first and a second part. The first part of the cost function depends on the first set of training samples, and the second part of the cost function depends on a first subset of training samples, the first subset being a subset of the second set of training samples. In addition, the second part of the cost function depends on an upper bound for the average prediction performance of the neural network for the first subset of training samples and the second part of the cost function is configured for preventing that the average prediction performance for the first subset of training samples exceeds the upper bound.