Medical Image Segmentation Using PCA Shape Regularization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing neural network-based medical image segmentation methods, such as UNET, often produce anatomically incorrect or meaningless outputs due to insufficient incorporation of shape regularization constraints.

Innovation Solution

Embedding shape regularization constraints within the neural network using a loss function that incorporates Principal Component Analysis (PCA) to learn the shape and deformation modes of organs, ensuring anatomically meaningful and regularized segmentations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If UNET-based segmentation algorithm is used, then regularized results are produced, but anatomical correctness is insufficient

Engineering Contradiction:
Improveanatomical correctnessVSAvoidmodel complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the UNET segmentation algorithm with a PCA-based shape regularization model. The combined system integrates the segmentation network's ability to produce regularized results with the PCA model's anatomical shape constraints, creating a unified framework that leverages both approaches to achieve anatomically correct segmentations.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The PCA-based shape regularization model acts as an intermediary between the UNET segmentation algorithm and the final segmentation output. It mediates by imposing anatomical shape constraints on the segmentation results, guiding the network to produce outputs that conform to known anatomical structures without directly modifying the network architecture.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If shape regularization constraints are embedded in loss function, then segmentation accuracy is improved, but computational complexity increases

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The PCA-based shape model is pre-computed from training data before the segmentation process. The deformation modes and mean shape are calculated in advance and stored, so during segmentation, only the application of these pre-computed constraints is required, reducing real-time computational complexity while maintaining accuracy improvements.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces a regularization parameter that controls the strength of the shape constraints in the loss function. By adjusting this parameter, the system can balance between segmentation accuracy and computational complexity, allowing flexible optimization based on specific application requirements.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4302236B1Image processing
Publication Date: 2025.08.27 KONINKLIJKE PHILIPS NV
  • EP4302236B1 patent drawingFigure 1
  • EP4302236B1 patent drawingFigure 2
  • EP4302236B1 patent drawingFigure 3

AI summary

Proposed are concepts for training a neural network (100), NN, for medical image segmentation. Such concepts include embedding of a shape regularization constraintsinto learned weights of the NN,. Such shape constraints can impose that the output of the segmentation process is anatomically meaningful and further regularized. In this way, improved performance for medical image segmentation may be achieved by proposed embodiments.