Medical Imaging Backbone Training with Self-Supervised Pretraining

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The lack of sufficient publicly available medical domain data and the high subjectivity and workload in radiology image analysis lead to inefficiencies in training effective machine learning algorithms for medical imaging, necessitating a more efficient use of available data.

Innovation Solution

A method involving self-supervised learning to train a backbone neural network, followed by supervised learning to update the network, utilizing both labeled and unlabeled medical data within healthcare organizations, and employing federated learning to aggregate and update multiple neural networks while ensuring data privacy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional supervised learning is used to train neural networks for medical image analysis, then the model can achieve good performance on labeled data, but the requirement for large amounts of expert-annotated training data increases significantly

Engineering Contradiction:
Improvedetection accuracyVSAvoidamount of training data
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent applies self-supervised learning as a preliminary action before supervised fine-tuning. The backbone neural network is first pre-trained on large amounts of unlabeled medical images using self-supervised learning, which extracts general features without requiring expert annotations. This preliminary training phase enables the model to learn from abundant unlabeled data, reducing the subsequent need for labeled training data while maintaining detection accuracy.

Inventive Principle:
Principle #10Preliminary action

2Adaptability or versatility

If each healthcare organization trains its own foundation model using its patient population data, then the model can be tailored to local characteristics, but the variability in patient population limits the generalizability of the trained model

Engineering Contradiction:
Improvelocal adaptabilityVSAvoidmodel generalizability
Core Design Contradiction:
Adaptability or versatilityVSReliability

Solution Approach 1:

The patent segments the training process into two distinct phases: (1) organization-specific fine-tuning that adapts the pre-trained backbone network to local patient population characteristics, and (2) aggregation of multiple organization-specific models into a federated model. This segmentation allows each organization to maintain local adaptability while the federated aggregation across diverse populations improves overall model generalizability and reliability.

Inventive Principle:
Principle #1Segmentation

3Quantity of substance

If public medical domain data is used for training foundation models, then large-scale training can be achieved, but the lack of sufficient publicly available data with permissible licenses limits the training process

Engineering Contradiction:
Improveavailable training dataVSAvoiddata accessibility
Core Design Contradiction:
Quantity of substanceVSEase of manufacture

Solution Approach 1:

The patent introduces federated learning as an intermediary mechanism that enables training on distributed private data without requiring data to be centralized or publicly accessible. Each healthcare organization keeps its data locally, and only model parameters are shared and aggregated. This intermediary approach bypasses the licensing and accessibility barriers of public data while still enabling large-scale training across multiple organizations' private datasets.

Inventive Principle:
Principle #24Intermediary (Mediator)

4Measurement precision

If manual annotation of medical findings is performed to create labeled training data, then the quality of training data improves, but the time and resources required for expert annotation increase significantly

Engineering Contradiction:
Improvetraining data qualityVSAvoidannotation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies partial action by using self-supervised learning to extract sufficient features from unlabeled data for the majority of the training process. Instead of requiring complete manual annotation of all training data, the model learns general patterns from unlabeled images and only requires limited labeled data for fine-tuning. This partial use of labeled data dramatically reduces annotation time while maintaining detection precision.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4621597A1Method for training a backbone neural network, systems, and computer program product
Publication Date: 2025.09.24 SIEMENS HEALTHINEERS AG
  • EP4621597A1 patent drawingFigure 1~2
  • EP4621597A1 patent drawingFigure 3
  • EP4621597A1 patent drawingFigure 4

AI summary

A method for training a first backbone neural network (V), comprising: training (S400) the first backbone neural network (V) using a first medical data set (MIS-1) using self-supervised learning, the first medical data set (MIS-1) having a first modality; training (S402) a first downstream neural network (N1) comprising: applying (S402-2) the trained first backbone neural network (N1) to a second medical data set (MIS-2) to provide a first feature vector (FV), the second medical data set (MIS-2) having the first modality; training (S402-4) the first downstream neural network (N1) with the first feature vector (FV) as input data and labels (LA1) associated with the second medical data set (MIS-1); and updating (S404) the trained first backbone neural network (V) based on a supervised training signal (TS1) generated during the training of the first downstream neural network (N1). A trained backbone neural network is provided by, in a first step, applying self-supervised learning and, in a second step, using supervised learning to update (finetune) the (partially) trained backbone neural network. Thereby, the required amount of training data is reduced.