Machine Learning Model for Flexible Pathology Progression Prediction

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for predicting the progression of diabetic retinopathy (DR) are limited as they can only provide predictions at specific future times, requiring retraining of the machine learning model for each time point, which is inflexible and inefficient.

Innovation Solution

A method for training a machine learning model that uses pairs of consecutive examination images from a longitudinal dataset to predict pathology progression at any intermediate time within a given time interval, allowing for flexible prediction of DR progression.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a machine learning model is trained to predict pathology progression at a specific future time, then the prediction accuracy at that time is improved, but the model requires retraining for each different time point, increasing device complexity and loss of time

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel training complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies universality by training a single machine learning model to predict pathology progression at multiple different future time points simultaneously, rather than requiring separate models for each time point. The model takes as input an examination image and a target future time, and outputs the predicted pathology state at that specific time, making the model multi-functional across different prediction time horizons.

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

Solution Approach 2:

The patent introduces time as an additional input dimension to the machine learning model. Instead of predicting pathology progression for fixed time points, the model now accepts the target future time as an input parameter, adding a temporal dimension to the prediction space. This allows the model to generalize across different time points without retraining.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If a machine learning model is retrained for each future time point to get accurate predictions, then the prediction accuracy is improved, but the loss of time increases due to repeated training

Engineering Contradiction:
Improveprediction accuracyVSAvoidmodel retraining time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent resolves this contradiction by creating a universal model that serves multiple time points simultaneously. The model is trained once on a dataset containing examples of pathology progression at various future times, and then can predict progression at any of those times without requiring retraining, thus eliminating the time loss associated with repeated training while maintaining accuracy.

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

Solution Approach 2:

The patent applies preliminary action by pre-training the model on a comprehensive dataset that includes pathology progression patterns across multiple future time points. This preliminary training equips the model with the knowledge to handle different time horizons, so when deployment occurs, no additional retraining is needed for different prediction times.

Inventive Principle:
Principle #10Preliminary action

3Adaptability or versatility

If examination images are analyzed to predict pathology progression at any future time, then the adaptability is improved, but the device complexity increases due to the need for flexible time-based prediction

Engineering Contradiction:
Improveprediction time flexibilityVSAvoidsystem complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent achieves adaptability through a universal prediction model that can handle any future time point within its training range. The model accepts the target time as input and adjusts its prediction accordingly, providing flexible adaptability without requiring complex separate systems for different time points.

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

Solution Approach 2:

The patent uses parameter changes by making the target future time a variable input parameter to the model. Instead of having fixed prediction time points, the model dynamically adjusts its prediction based on the input time parameter, enabling flexible adaptation to different prediction horizons through simple parameter variation rather than structural complexity.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4535233A1Method and apparatus for predicting progression of a pathology
Publication Date: 2025.04.09 UNIVERSITE DE BRETAGNE OCCIDENTALE (UBO)
  • EP4535233A1 patent drawingFigure 1~3
  • EP4535233A1 patent drawingFigure 4A
  • EP4535233A1 patent drawingFigure 4B

AI summary

A method and apparatus are proposed for training a model for predicting information representative of a pathology progression at a future time (t+Δt) from one or more examination images (for instance one image 2D + a 3D volume) of a region of interest of a patient captured at a current time (t), said method comprising: - getting (20) a training dataset, comprising time series of examination images of said region of interest from patients associated with labels, said labels (S(t)) comprising said information representative of a pathology progression, - training (E1, 21, 22) the machine learning model (MLM) with inputs from said training dataset comprising, at least one of said inputs comprising a pair of consecutive examination images comprising a first image (x(ti)) associated with a first time (ti) and first label (S(ti)), a second image (x(ti+1)) associated with a second time (ti+1) and a second label (S(ti+1)), to output a target intermediate label (TS(tmix)) at an intermediate time selected within a time interval ([ti,ti+1]) between the first and second examination images, said target intermediate label (TS(tmix)) being determined for said intermediate time (tmix), based on a given pathology progression profile (l(t))