Image Segmentation via Latent Variable Sampling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning models for image segmentation often fail to capture the inherent ambiguities in medical imaging, leading to potential misdiagnosis and sub-optimal treatment by providing only the most likely segmentation, which may not account for complex variations and ambiguities in lesion boundaries and malignancy.

Innovation Solution

A segmentation system that generates multiple plausible segmentations of an image by sampling latent variables from a hierarchical latent space, allowing for the modeling of diverse and varying segmentation structures across scales, thereby addressing the ambiguities in medical imaging.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a single machine learning model generates only the most likely segmentation, then the processing speed is fast, but the reliability of diagnosis is reduced due to inability to capture ambiguities

Engineering Contradiction:
Improvediagnostic reliabilityVSAvoidprocessing speed
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent segments the segmentation process by generating multiple distinct segmentation hypotheses (e.g., 5-10 different segmentations) rather than producing a single result. Each segmentation represents a different plausible interpretation of ambiguous image regions, allowing doctors to review multiple possibilities and improve diagnostic reliability without requiring complete reprocessing of the entire image for each hypothesis.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system dynamically adjusts the number and characteristics of generated segmentations based on image complexity and ambiguity levels. The model can generate a smaller number of segmentations for simple images and more segmentations for ambiguous cases, optimizing the balance between processing speed and diagnostic reliability for each specific input.

Inventive Principle:
Principle #15Dynamics

2Adaptability or versatility

If multiple different systems are used to generate multiple segmentations, then the diversity of segmentations is improved, but the device complexity and resource consumption increase

Engineering Contradiction:
Improvesegmentation diversityVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent employs a single multi-functional machine learning model that can generate multiple diverse segmentations from one image input. This universal model replaces the need for multiple separate systems, achieving segmentation diversity through internal model capabilities rather than external system multiplication, thereby reducing overall device complexity and resource consumption.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system uses a hierarchical latent space structure where multiple segmentation hypotheses are nested within a single model framework. The model processes the image once to extract features, then generates multiple segmentations by sampling from latent variables, effectively nesting multiple output possibilities within a single processing pipeline rather than requiring multiple separate processing chains.

Inventive Principle:
Principle #7Nested doll (Nesting)

3Measurement precision

If manual segmentations are performed by multiple people, then the accuracy and coverage of segmentations is improved, but the time consumption and resource usage increase

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidtime consumption
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The machine learning model performs self-service by automatically generating multiple diverse segmentations without requiring human intervention for each segmentation. The model autonomously explores different segmentation hypotheses by sampling from its latent space, achieving the accuracy and coverage that would otherwise require multiple manual segmentations while eliminating the time consumption associated with human review and coordination.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system performs preliminary computation by generating multiple segmentation hypotheses in advance before final diagnostic decisions are made. This preliminary generation of diverse segmentations allows the system to prepare multiple analysis paths simultaneously, reducing the time that would otherwise be required for iterative manual review and adjustment by multiple professionals.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3973508B1Sampling latent variables to generate multiple segmentations of an image
Publication Date: 2025.01.01 DEEPMIND TECH LTD
  • EP3973508B1 patent drawingFigure 1
  • EP3973508B1 patent drawingFigure 2Aa
  • EP3973508B1 patent drawingFigure 2Bb

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a plurality of possible segmentations of an image. In one aspect, a method comprises: receiving a request to generate a plurality of possible segmentations of an image; sampling a plurality of latent variables from a latent space, wherein each latent variable is sampled from the latent space in accordance with a respective probability distribution over the latent space that is determined based on the image; generating a plurality of possible segmentations of the image, comprising, for each latent variable, processing the image and the latent variable using a segmentation neural network having a plurality of segmentation neural network parameters to generate the possible segmentation of the image; and providing the plurality of possible segmentations of the image in response to the request.