System and method for synthesizing images containing pathologies for training discriminative prediction algorithm

The generative AI framework synthesizes realistic pathological images to address data imbalance and annotation costs in AI training, enhancing model reliability and accuracy for medical imaging.

WO2026131217A1PCT designated stage Publication Date: 2026-06-25KONINKLIJKE PHILIPS NV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/EP2025/085806
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-04-29
Filing Date
2025-12-08
Publication Date
2026-06-25

AI Technical Summary

Technical Problem

Existing artificial intelligence (AI) systems for medical imaging, particularly ultrasound, face challenges in training due to data imbalance, high annotation costs, and reduced accuracy in underrepresented subgroups, especially for rare diseases, lacking quality assurance mechanisms.

Method used

A method and system using a generative AI framework to synthesize realistic pathological images by pairing negative images with binary mask images, employing a pathology-anatomy bank and conditional diffusion models to ensure accurate placement and structure of pathologies, increasing data diversity and prevalence.

Benefits of technology

Enhances training data diversity, improves AI model reliability, and provides a quality control mechanism for synthesized images, ensuring accurate pathology detection and diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure EP2025085806_25062026_PF_FP_ABST
    Figure EP2025085806_25062026_PF_FP_ABST
Patent Text Reader

Abstract

A method is provided for synthesizing images containing pathologies for training a discriminative prediction algorithm for performing pathology detection and diagnosis. The method includes randomly selecting a negative image showing no pathology from a dataset of negative images; identifying a location of an anatomical landmark in the negative image; determining a desired location for placing a pathology to be simulated in the negative image based on the anatomical landmark; generating a binary mask image based on the desired location for the pathology, the binary mask image including a basic structure of the pathology to be simulated at the desired location; selecting an image pair including the negative image with a masked out area and the binary mask image; and generating a synthesized image based on the image pair, where the synthesized image includes a synthesized background corresponding to the negative image and a synthesized foreground corresponding to the basic structure of the pathology to be simulated in the binary mask image at the masked out area in the negative image.
Need to check novelty before this filing date? Find Prior Art