GAN Training Loop for Real-Time Medical Image False-Positive Filtering

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing object detection systems in medical imaging, such as endoscopy, struggle with inaccurate feature detection, reliance on manually annotated training sets, difficulty in differentiating false positives from true positives, and delays in real-time analysis, leading to inefficiencies and potential misdiagnosis.

Innovation Solution

A computer-implemented system using generative adversarial networks (GANs) for training neural networks to enhance object detection, incorporating a two-phase training loop with a perception branch for initial detection and a generative-adversarial branch for refining detections, enabling real-time differentiation between true and false positives.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If conventional training methods are used for GANs, then training time and computational resources are excessive, but the quality and realism of generated images remain insufficient

Engineering Contradiction:
Improvetraining speedVSAvoidimage quality
Core Design Contradiction:
ProductivityVSManufacturing precision

Solution Approach 1:

The patent introduces a pre-trained autoencoder as an intermediary component that compresses input images into latent representations before feeding them to the GAN. This mediator enables faster training by reducing the dimensionality and complexity of the data the GAN must process, while the perceptual loss function using VGG network ensures high-quality output by comparing feature representations rather than raw pixels

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent replaces the traditional pixel-based mean squared error loss function with a perceptual loss function that operates in feature space using a pre-trained VGG network. This substitution changes the training mechanism from direct pixel comparison to semantic feature comparison, achieving better image quality and faster convergence by operating in a more meaningful representation space

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Manufacturing precision

If extensive training data and computational resources are allocated, then generated images achieve higher realism, but training time and resource consumption increase significantly

Engineering Contradiction:
Improveimage realismVSAvoidtraining time
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent performs preliminary actions by pre-training the autoencoder separately before GAN training, and by using a pre-trained VGG network for perceptual loss calculation. These preliminary preparations enable the main GAN training to converge faster with fewer iterations, reducing overall training time while maintaining high image realism through the established feature extraction capabilities

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent extracts the feature extraction functionality into a separate pre-trained VGG network that is used solely for calculating perceptual loss. This extraction allows the GAN to focus on generating realistic images while the VGG network handles the complex task of semantic feature comparison, improving training efficiency without sacrificing image realism

Inventive Principle:
Principle #2Taking out (Extraction)

3Manufacturing precision

If traditional loss functions are used for training GANs, then training is simpler to implement, but generated images lack diversity and realism

Engineering Contradiction:
Improveimage qualityVSAvoidtraining complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple loss components (reconstruction loss from autoencoder, adversarial loss from GAN, and perceptual loss from VGG network) into a unified training framework. This combination leverages the strengths of each component: the autoencoder ensures data fidelity, the GAN provides realism, and the perceptual loss maintains semantic quality, achieving superior image quality while the modular structure keeps implementation manageable

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP3762868B1Systems and methods for training generative adversarial networks and use of trained generative adversarial networks
Publication Date: 2026.05.06 COSMO ARTIFICIAL INTELLIGENCE AI LTD
  • EP3762868B1 patent drawingFigure 1
  • EP3762868B1 patent drawingFigure 2
  • EP3762868B1 patent drawingFigure 3

AI summary

The present disclosure relates to computer-implemented systems and methods for training and using generative adversarial networks. In one implementation, a system for training a generative adversarial network may include at least one processor that may provide a first plurality of images including representations of a feature-of-interest and indicators of locations of the feature-of- interest and use the first plurality and indicators to train an object detection network. Further, the processor(s) may provide a second plurality of images including representations of the feature-of-interest, and apply the trained object detection network to the second plurality to produce a plurality of detections of the feature-of- interest. Additionally, the processor(s) may provide manually set verifications of true positives and false positives with respect to the plurality of detections, use the verifications to train a generative adversarial network, and retrain the generative adversarial network using at least one further set of images, further detections, and further manually set verifications.