Domain Adaptation for Prostate MRI Cancer Detection Drift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI-based systems for prostate cancer detection using mpMRI imaging suffer from decreased performance when encountering out-of-distribution image acquisition parameters, leading to overconfident predictions.
Innovation Solution
A domain adaptation framework that generates synthesized medical images with in-distribution properties from out-of-distribution images using a machine learning-based generator network, trained with a discriminator network to maintain semantic accuracy without additional annotations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI-based systems are trained with mpMRI imaging acquired with specific image acquisition parameters, then detection performance is improved for in-distribution images, but performance significantly decreases for out-of-distribution images acquired with different image acquisition parameters
Solution Approach 1:
The patent introduces a domain adaptation module as an intermediary between the input image and the detection model. This module transforms out-of-distribution images (acquired with different parameters) into in-distribution images (matching training data characteristics) before they reach the detection model, thereby maintaining reliable performance across varying acquisition conditions without retraining the entire system
Solution Approach 2:
The domain adaptation module dynamically adjusts image parameters through transformation operations to convert out-of-distribution images into in-distribution images. By changing the parameter distribution of input images to match training data characteristics, the system maintains consistent detection performance across different MRI acquisition settings
2Productivity
If AI-based systems are trained on limited distribution data, then training efficiency is improved, but the system produces overconfident predictions on out-of-distribution data
Solution Approach 1:
The system performs domain adaptation as a preliminary action before the main detection task. By pre-processing out-of-distribution images to match training data characteristics, the system prepares the input data in advance, ensuring that the detection model receives appropriately distributed inputs without requiring extensive retraining or additional annotated data
Data Source
AI summary
Systems and methods for performing a medical imaging analysis task using a machine learning based model are provided. One or more input medical images acquired using one or more out-of-distribution image acquisition parameters and having out-of-distribution imaging properties are received. The one or more out-of-distribution image acquisition parameters and the out-of-distribution imaging properties are out-of-distribution with respect to training data on which the machine learning based model is trained. One or more synthesized medical images are generated from the one or more input medical images using a machine learning based generator network. The one or more synthesized medical images are generated for one or more in-distribution image acquisition parameters and have in-distribution imaging properties. The one or more in-distribution image acquisition parameters and the in-distribution imaging properties are in-distribution with respect to the training data on which the machine learning based model is trained. The medical imaging analysis task is performed based on the one or more synthesized medical images using the machine learning based model. Results of the medical imaging analysis task are output.