3D Tumor Risk Mapping With Mixed Supervision for Sparse Labels
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine-learning systems for tumor identification and grading in medical imaging face challenges due to the lack of strong groundtruth data, particularly in medical imaging where semantic labeling is sparse, leading to low accuracy in detecting clinically-significant cancer and missing significant cancer outside visible lesions.
Innovation Solution
A method that incorporates multiple sources of pathology data, including lesion and systematic biopsies, to enhance cancer grading by integrating routinely collected clinical data, using a combination of strong and weak supervision techniques to improve the granularity and accuracy of cancer detection and grading in a 3D space.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fully supervised algorithms are used for semantic segmentation in medical imaging, then segmentation accuracy is improved, but the requirement for annotated training data increases significantly
Solution Approach 1:
The patent segments the supervision process into multiple levels: fully supervised learning on a small subset of annotated data to learn precise tumor boundaries, and weakly supervised learning on larger unannotated datasets to capture general tumor patterns. This hierarchical segmentation of supervision strength allows the system to achieve high accuracy while reducing dependency on large amounts of annotated data.
Solution Approach 2:
The patent performs preliminary annotation on a small subset of data to create high-quality training examples. These pre-annotated samples are used to initialize the model with strong supervision, establishing a foundation of accurate tumor boundary detection before expanding to larger datasets with weaker supervision signals.
2Measurement precision
If more annotated data is collected to improve model performance, then detection accuracy is improved, but the time and cost for data annotation increases
Solution Approach 1:
The patent applies partial supervision by using fully annotated data for a small portion of the training process and weakly supervised signals for the majority of training. This partial application of expensive annotation resources achieves most of the potential accuracy gains while avoiding the prohibitive costs of annotating entire datasets.
Solution Approach 2:
The patent introduces an intermediary weak supervision mechanism that bridges the gap between fully annotated and completely unannotated data. This intermediary layer uses approximate labels and relaxed constraints to guide learning on large datasets without requiring precise manual annotations, thereby reducing annotation time while maintaining detection accuracy.
3Measurement precision
If conventional semantic segmentation is used, then pixel-level classification is achieved, but the ability to handle sparse and ambiguous boundaries in medical images deteriorates
Solution Approach 1:
The patent employs dynamic supervision strategies that adapt to local image characteristics. In regions with clear boundaries, fully supervised pixel-level classification is applied. In regions with ambiguous or sparse boundaries, the system dynamically switches to weakly supervised approaches that are more robust to labeling uncertainty, allowing flexible adaptation to varying boundary difficulties across different regions of the medical images.
Data Source
AI summary
The present disclosure relates to techniques for non-invasive tumor identification, classification, and grading using mixed exam-, region-, and voxel-wise supervision. Particularly, aspects are directed to a computer implemented method that includes obtaining medical images of a subject, inputting the medical images into a three-dimensional neural network model constructed to produce a voxelwise cancer risk map of lesion occupancy and cancer grade as two output channels using an objective function having a first loss function that captures strongly supervised loss for regression in lesions and a second loss function that captures weakly supervised loss for regression in regions, generating an estimated segmentation boundary around one or more lesions, predicting a cancer grade for each pixel or voxel within the medical images, and outputting the voxelwise cancer risk map of lesion occupancy determined based on the estimated segmentation boundary and the cancer grade for each pixel or voxel within the medical images.