Patient-Specific Auto-Segmentation Through Online Model Adaptation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing static machine learning models for medical image segmentation in radiotherapy are not patient-specific, leading to inaccuracies due to variations in patient-specific factors and physician preferences, requiring repeated manual corrections in adaptive radiotherapy.
Innovation Solution
Adaptive training of segmentation models 'online' using physician corrections as ground-truth data to iteratively fine-tune the models, incorporating patient-specific and physician-specific adjustments over multiple treatment sessions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If a static machine learning model is used for segmentation, then the model structure is simple and easy to deploy, but the segmentation accuracy deteriorates due to inability to adapt to patient-specific factors and physician preferences
Solution Approach 1:
The patent transforms the static segmentation model into a dynamic system that adapts to individual patients through online learning. The model evolves from a fixed structure to one that continuously updates its parameters based on physician corrections and patient-specific data, resolving the contradiction between model simplicity and segmentation accuracy.
Solution Approach 2:
The system changes the parameters of the segmentation model during deployment by incorporating physician corrections as ground-truth data to fine-tune the model for each patient. This parameter adaptation allows the model to maintain simplicity while achieving high segmentation accuracy specific to each patient's anatomy and physician preferences.
2Speed
If a static machine learning model is used for segmentation, then the deployment process is fast, but the time required for manual corrections increases due to repeated adjustments needed for each patient
Solution Approach 1:
The patent implements a feedback loop where physician corrections to segmentation results are captured and used to fine-tune the model for subsequent patients. This feedback mechanism reduces the manual correction time for each new patient while maintaining fast deployment, as the model learns from previous corrections rather than starting from scratch.
Solution Approach 2:
The system performs preliminary adaptation by fine-tuning the segmentation model using physician corrections from previous patients before deploying it to new patients. This preliminary action reduces the manual correction time required for each patient while preserving the fast deployment capability of the static model structure.
3Measurement precision
If physician corrections are manually applied to each segmentation, then the segmentation accuracy improves for individual patients, but the overall workflow efficiency deteriorates due to repeated manual interventions
Solution Approach 1:
The patent enables the segmentation model to serve itself by automatically learning from physician corrections. Instead of requiring physicians to manually correct each segmentation, the system autonomously adapts to patient-specific factors and physician preferences, improving accuracy while maintaining workflow efficiency.
Solution Approach 2:
The patent replaces the mechanical process of repeated manual corrections with an automated online learning mechanism. The system substitutes the manual intervention workflow with an algorithmic adaptation process that automatically fine-tunes the model using physician feedback, thereby improving accuracy without sacrificing productivity.
Data Source
AI summary
Disclosed herein are systems and methods for adaptively training machine learning models for auto-segmentation of medical images. A system executes a segmentation model that receives a medical image as input and generates an initial segmentation of the medical image for a radiotherapy treatment. The system identifies a corrected segmentation corresponding to the initial segmentation generated in response to an input at a user interface presenting the initial segmentation. The system fine-tunes the segmentation model based on the medical image and the corrected segmentation to generate a fine-tuned segmentation model.