Unsupervised ML Model for 3D Object Analysis from 2D Images
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning models for analyzing 3D objects from 2D images captured at different views require supervised training and manual annotation, which is time-consuming and expensive, and often result in inaccurate detection of visual findings like breast cancer due to the lack of sufficient correctly annotated images and reliance on heuristic approaches that confuse close candidates.
Innovation Solution
An unsupervised machine learning model training method that divides 3D images into patches, computes corresponding 2D images from different orientations, automatically labels patch pairs, and creates a training dataset for patch measure computation, enabling the model to output likelihood of visual findings without external annotation, specifically for breast cancer detection in mammography.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If supervised training with manual annotation is used, then the ML model can be trained, but the process is time-consuming and expensive
Solution Approach 1:
The system performs self-annotation by automatically generating labels through geometric correspondence between 2D patches and 3D regions. The ML model trains on self-generated data without requiring external human annotation, thus eliminating the time-consuming manual labeling process while maintaining training capability
Solution Approach 2:
The system pre-computes 3D images from 2D mammographic images and pre-establishes the geometric mapping between 2D patches and 3D regions before training. This preliminary preparation creates ready-to-use training data with automatic labels, significantly reducing the time required during the actual training phase
2Measurement precision
If supervised training with manual annotation is used, then the ML model can be trained, but the cost is high
Solution Approach 1:
The system eliminates the need for expensive human annotators by implementing self-annotation through automatic geometric correspondence. The training data is labeled programmatically using the established mapping between 2D patches and 3D regions, drastically reducing annotation costs
Solution Approach 2:
The system creates synthetic training data by generating 3D images from existing 2D mammographic images and creating corresponding patch pairs. This copying and transformation process generates sufficient training examples without requiring expensive manual creation of annotated datasets
3Productivity
If heuristic approaches are used, then the system can identify candidates, but close candidates are confused
Solution Approach 1:
The system divides the image space into discrete patches and processes them individually. By segmenting the analysis into distinct 2D patches from different views, the system can evaluate each region separately and reduce confusion between close candidates that appear in different patch locations
Solution Approach 2:
The system transitions from 2D patch analysis to 3D region analysis by establishing geometric correspondence. Candidates are evaluated in the additional 3D dimension, allowing the system to distinguish between close candidates that may appear similar in 2D but occupy different 3D spaces, thereby improving detection accuracy
4Quantity of substance
If manual annotation is required, then training data can be created, but sufficient correctly annotated images are lacking
Solution Approach 1:
The system generates large quantities of training data by copying and transforming existing 2D mammographic images into 3D representations and creating multiple 2D patch pairs from each 3D image. This multiplicative data generation process creates sufficient training examples without requiring proportional increases in manual annotation
Solution Approach 2:
The system creates accurately labeled training data through self-annotation using geometric correspondence. The automatic labeling process ensures high annotation accuracy by relying on precise geometric relationships rather than human interpretation, eliminating the trade-off between quantity and accuracy
Data Source
AI summary
There is provided a method for unsupervised training of a machine learning model, comprising: receiving 3D images depicting a respective object, for each respective 3D image: dividing the 3D image into 3D patches, computing a first 2D image corresponding to a first orientation of the respective object, computing a second 2D image corresponding to a second orientation, automatically labelling pairs of 2D patches from the first and second 2D images with a patch measure indicative of likelihood of a certain 3D patch of the 3D image corresponding to a certain pair of 2D patches, training the ML model using a training dataset including the labelled patch pairs, for receiving patches extracted from first and second 2D images captured by an imaging sensor at the first and second orientations, and outputting an indication of likelihood of a visual finding in a 3D region of the object corresponding to the 2D patches.