Shift Invariant Loss for Deep Learning Image Segmentation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The accuracy of convolutional neural networks in feature detection tasks, such as image segmentation, is significantly affected by misalignment between input images and labeled images, which can be time-consuming to correct manually and requires additional training processes when using separate alignment networks.

Innovation Solution

A method that aligns predicted label images with true labeled images by computing a similarity measure using techniques like cross-correlation or mutual information, allowing for automatic adjustment of transformations such as shifts and rotations to optimize neural network loss during training.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If manual adjustment of input images and label images is performed to solve misalignment, then alignment accuracy is improved, but training time and operational complexity increase significantly

Engineering Contradiction:
Improvealignment accuracyVSAvoidtraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs self-alignment by automatically computing transformations between input images and label images without requiring manual intervention. The neural network itself generates the alignment transformations, allowing the system to correct its own misalignment issues during training.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system changes the parameter space by introducing transformation parameters (such as spatial transformations) that are learned and optimized during training. Instead of fixing alignment manually, the system dynamically adjusts transformation parameters to achieve proper alignment between inputs and labels.

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If a separate neural network is used to align input and labels, then alignment accuracy is improved, but system complexity and training resources increase

Engineering Contradiction:
Improvealignment accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The alignment functionality is merged into the main neural network training process. Instead of using a separate alignment network, the system integrates transformation computation and optimization directly into the segmentation network's training loop, combining multiple functions into a single unified system.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The main neural network is given multi-functionality by enabling it to both perform segmentation and compute alignment transformations. This universal approach allows the same network to handle both alignment and segmentation tasks, eliminating the need for specialized separate networks.

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

3Measurement precision

If a separate neural network is used to align input and labels, then alignment accuracy is improved, but training computational load increases

Engineering Contradiction:
Improvealignment accuracyVSAvoidcomputational load
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The alignment computation is merged with the segmentation training process, so that both alignment parameters and segmentation weights are optimized simultaneously in a single training loop. This combines computational tasks that would otherwise run separately, improving efficiency.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The alignment optimization continues continuously throughout the training process rather than requiring a separate pre-training phase. The useful action of alignment computation is performed continuously alongside segmentation learning, maximizing computational efficiency throughout the entire training duration.

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP4091093B1Shift invariant loss for deep learning based image segmentation
Publication Date: 2025.04.16 VERILY LIFE SCIENCES LLC
  • EP4091093B1 patent drawingFigure 1
  • EP4091093B1 patent drawingFigure 2
  • EP4091093B1 patent drawingFigure 3

AI summary

Systems and methods of improving alignment in dense prediction neural networks are disclosed. A method includes identifying, at a computing system, an input data set and a label data set with one or more first parts of the input data set corresponding to a label. The computing system processes the input data set using a neural network to generate a predicted label data set that identifies one or more second parts of the input data set predicted to correspond to the label. The computing system determines an alignment result using the predicted label data set and the label data set and a transformation of the one or more first parts, including a shift, rotation, scaling, and/or deformation, based on the alignment result. The computing system computes a loss score using the transformation, label data and the predicted label data set and updates the neural network based on the loss score.