Mask-Based Medical Imaging for Diverse Synthetic Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The development of AI and ML-enabled medical devices faces challenges in acquiring large, diverse datasets for training due to stringent medical environments, leading to potential misdiagnosis issues.
Innovation Solution
A system and method for generating conditioned masks from an initial mask applied to a master image, adjusting dimensions and parameters to create a series of frames, which simulate various anatomical and imaging conditions, enabling the synthesis of diverse medical images and videos.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI/ML models are trained with real clinical data, then the models can achieve high diagnostic accuracy, but acquiring sufficient diverse data is difficult due to medically stringent environments
Solution Approach 1:
The patent creates synthetic medical images by copying and modifying real image structures through mask-based generation. The system generates synthetic images that replicate anatomical variations, pathologies, and imaging conditions without requiring actual patient data, thus achieving data diversity while maintaining diagnostic accuracy for AI model training.
Solution Approach 2:
The system applies multiple mask parameters (size, shape, position, orientation) to modify the initial mask and generate varied synthetic images. By systematically changing these parameters, the system creates diverse training datasets that cover various anatomical variations, pathologies, and imaging conditions, resolving the contradiction between data diversity and acquisition difficulty.
2Ease of manufacture
If a single mask is applied to generate synthetic images, then the generation process is simple, but the images lack diversity for robust model training
Solution Approach 1:
The patent segments the mask application process into multiple independent mask parameters (size parameter, shape parameter, position parameter, orientation parameter). Each parameter can be independently adjusted to generate different variations of synthetic images, maintaining simplicity in the base process while achieving high diversity through parameter combination.
Solution Approach 2:
The system transitions from a static single mask to dynamic mask generation by varying parameters such as size, shape, position, and orientation. This dynamic approach allows the same base mask to generate multiple diverse images by adjusting parameters, achieving both simplicity in the base process and diversity in the output.
3Reliability
If extensive labeled datasets are collected, then AI models can be trained robustly, but the process is time-consuming and resource-intensive
Solution Approach 1:
The system performs preliminary actions by generating synthetic images before actual clinical data collection. By pre-generating diverse synthetic datasets covering various anatomical variations, pathologies, and imaging conditions, the system reduces or eliminates the need for time-consuming real data collection and labeling processes while maintaining training robustness.
Solution Approach 2:
The system serves itself by generating its own training data through mask-based synthetic image generation. Instead of relying on external data collection efforts, the system creates self-contained synthetic datasets that are automatically labeled and ready for training, eliminating the time-consuming human annotation process while ensuring data diversity for robust model training.
Data Source
AI summary
Various systems and methods are presented regarding synthesizing images for application with a medical imaging system. An initial mask can be generated from one or more regions of interest (RoI) on an initial image. The one or more RoIs in the initial mask can undergo modification to create a series of conditioned masks which can be subsequently applied to the initial image to create a respective image(s) modified in accordance with the modified RoIs in the respective conditioned mask. The series of respective images can be used as frames in a cineloop. Hence, a sequence of modified images/frames can be generated from a single initial image and mask. The conditioned masks can be modified to reflect a range of healthy and unhealthy conditions. The frames/images can be utilized to train medical imaging models.