Medical Image Saliency Maps With Shape-Constrained AI Explanations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Neural network-based medical image classification systems lack transparency, providing unclear explanations for their predictions, which reduces the value and trustworthiness for clinicians.

Innovation Solution

An explainable medical image classification system that generates saliency maps alongside classifications, ensuring accurate and transparent predictions by training with dual loss functions to improve both classification and explanation quality.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a neural network model is used for medical image classification, then classification accuracy is improved, but interpretability and explanation quality deteriorate

Engineering Contradiction:
Improveclassification accuracyVSAvoidexplanation quality
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent introduces saliency maps as an intermediary component that bridges the neural network model and the clinician. The saliency map generator creates visual explanations highlighting important image regions, serving as a mediator that translates the model's internal decision-making process into interpretable visual feedback without affecting the model's classification accuracy

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent implements a feedback mechanism where saliency maps are generated based on the neural network's classification process and fed back to the clinician. This feedback loop provides interpretability by showing which image regions influenced the classification, allowing clinicians to verify and understand the model's reasoning while maintaining high classification accuracy

Inventive Principle:
Principle #23Feedback

2Loss of information

If saliency maps are generated for all classifications, then explanation quality is improved, but reliability of results deteriorates due to low-quality explanations

Engineering Contradiction:
Improveexplanation qualityVSAvoidresult reliability
Core Design Contradiction:
Loss of informationVSReliability

Solution Approach 1:

The patent changes the parameter of saliency map generation from a binary on/off decision to a confidence-based continuous parameter. By using confidence thresholds to control when saliency maps are generated and displayed, the system adapts the explanation quality parameter based on the reliability of the classification, ensuring that only high-confidence classifications with meaningful explanations are presented to clinicians

Inventive Principle:
Principle #35Parameter changes

3Reliability

If confidence thresholds are applied to filter low-confidence classifications, then result reliability is improved, but productivity deteriorates due to additional filtering steps

Engineering Contradiction:
Improveresult reliabilityVSAvoiddiagnostic efficiency
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The patent applies confidence threshold filtering as a preliminary action during the classification process itself, rather than as a separate post-processing step. The neural network model evaluates confidence levels and automatically filters or flags low-confidence classifications before they reach the clinician, eliminating the need for manual review and maintaining diagnostic efficiency while improving result reliability

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4685735A1Enhanced interpretability of ai results in medical imaging by supervision, and associated devices, sytems, and methods
Publication Date: 2026.01.28 KONINKLIJKE PHILIPS NV
  • EP4685735A1 patent drawingFigure 1
  • EP4685735A1 patent drawingFigure 2
  • EP4685735A1 patent drawingFigure 3

AI summary

An explainable medical image classification system receives a medical image. The system generates, by a trained predictive network, a classification representative of a patient physiology in the medical image. The system generates a saliency map based on the medical image and the trained predictive network. The system compares features of the saliency map to a pre-determined shape constraint. The system outputs the saliency map, the medical image, and the classification, when the features of the saliency map satisfy the pre-determined shape constraint. The system prevents output of the saliency map when the saliency map does not satisfy the pre-determined shape constraint.