Cerebral Hemorrhage CT Detection Using Size-Adaptive Neural Decoding

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing deep learning-based image recognizers for cerebral hemorrhage diagnosis suffer from low detection accuracy and struggle to handle hemorrhages of varying sizes and shapes in brain CT images, often misrecognizing normal areas as bleeding areas or failing to detect small hemorrhages.

Innovation Solution

A deep neural network-based system using a U-Net model with a center surround difference method to detect and classify bleeding areas and suspected areas by size, applying neural networks of varying depths to enhance accuracy and adaptability.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a single deep learning model is used for cerebral hemorrhage detection, then the system structure is simple, but the detection accuracy is low and cannot handle hemorrhages of various sizes effectively

Engineering Contradiction:
Improvedetection accuracyVSAvoidsystem structure
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system divides the detection task into multiple segments by creating separate neural network models for different hemorrhage size categories (small, medium, large). Each model is specialized for detecting hemorrhages within its specific size range, which improves detection accuracy for each category while managing system complexity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system dynamically selects which neural network model to apply based on the size of the detected hemorrhage area. This dynamic adaptation allows the system to optimize detection accuracy for each specific case by choosing the most appropriate model, rather than using a fixed single-model approach.

Inventive Principle:
Principle #15Dynamics

2Measurement precision

If doctors manually diagnose cerebral hemorrhage from CT images, then flexibility in handling various cases is maintained, but detection accuracy varies and time-consuming

Engineering Contradiction:
Improvedetection accuracyVSAvoiddiagnosis time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary automated detection and classification of hemorrhage areas before final diagnosis confirmation. By pre-identifying and categorizing hemorrhage regions using the multi-model neural network system, it reduces the time required for manual review while maintaining high detection accuracy, allowing doctors to focus on confirming results rather than initial detection.

Inventive Principle:
Principle #10Preliminary action

3Reliability

If normal areas are distinguished from bleeding areas, then false positives are reduced, but small bleeding areas may be missed

Engineering Contradiction:
Improvediagnosis reliabilityVSAvoiddetection accuracy
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The system applies different detection strategies and neural network models to different local regions based on hemorrhage size characteristics. Small hemorrhage areas are processed with models optimized for fine-detail detection, while larger areas use models suited for broader pattern recognition. This local adaptation improves both reliability in distinguishing normal from abnormal areas and precision in detecting small hemorrhages.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4230142B1Deep neural network-based cerebral hemorrhage diagnosis system
Publication Date: 2026.02.11 KYUNGPOOK NAT UNIV IND ACADEMIC COOP FOUND
  • EP4230142B1 patent drawingFigure 1
  • EP4230142B1 patent drawingFigure 2
  • EP4230142B1 patent drawing

AI summary

The present invention relates to a deep neural network-based cerebral hemorrhage diagnosis system comprising: an input unit which receives a CT image and presents bleeding areas and suspected bleeding areas; a bleeding size classification unit which is provided with the bleeding areas and suspected bleeding areas presented by the input unit and classifies the bleeding areas and suspected bleeding areas by size; a decoding unit which decodes the bleeding areas and suspected bleeding areas, classified by size by the bleeding size classification unit, by applying neural networks of different depth according to the size of the bleeding areas and suspected bleeding areas; and an output unit which sums and outputs, as a final bleeding area, the results decoded according to size by the decoding unit.