How to generate rare medical images for training deep learning algorithms

By combining majority and minority segmentation masks to generate synthetic medical images, the method addresses the challenge of rare disease data representation, enhancing the training of deep learning algorithms for anatomical abnormality detection and characterization.

JP7874102B2Active Publication Date: 2026-06-15QUANTUM SURGICAL
View PDF 4 Cites 0 Cited by

Patent Information

Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
QUANTUM SURGICAL
Filing Date
2022-05-06
Publication Date
2026-06-15

AI Technical Summary

⚠Technical Problem

Existing methods for training deep learning algorithms to detect or characterize anatomical abnormalities in medical images face challenges due to the rarity of diseases, leading to insufficient data representation and potential overfitting, particularly when using synthetic images generated by generative adversarial networks.

⚗Method used

A method involving the generation of synthetic medical images using a combination of majority and minority segmentation masks, where majority masks represent normal tissues and minority masks represent abnormal tissues, utilizing a neural network to create diverse and realistic synthetic images, thereby increasing the variety and balance of training data without requiring multiple generative adversarial networks.

🎯Benefits of technology

This approach generates a diverse set of synthetic images that effectively train machine learning algorithms, reducing overfitting and improving the detection or characterization of abnormalities by ensuring a balanced representation of normal and abnormal tissues, thus enhancing the algorithm's generalization capability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007874102000001
    Figure 0007874102000001
  • Figure 0007874102000002
    Figure 0007874102000002
  • Figure 0007874102000003
    Figure 0007874102000003
Patent Text Reader

Abstract

The present invention relates to a method (100) for generating a synthetic medical image representative of a biological tissue of interest and an anomaly within said biological tissue of interest. The method (100) includes generating (101) a majority segmentation mask associated with a real medical image without the anomaly, generating (102) a minority segmentation mask associated with the real medical image with the anomaly, training a neural network to generate a synthetic medical image based on the segmentation masks, generating (103) an artificial segmentation mask based on the majority and minority segmentation masks by combining the segmentation of the anomaly with the minority segmentation mask and the segmentation of the biological tissue with the majority segmentation mask, and generating (105) a synthetic medical image based on the artificial segmentation mask and by using the previously trained neural network.
Need to check novelty before this filing date? Find Prior Art