Constrained Deep Learning Network for Interpretable Biomarker Extraction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current medical imaging technologies face challenges in generating interpretable and data-driven biomarkers, as deep learning systems lack interpretability and require large datasets, while clinical studies rely on hand-crafted features that are not reproducible or standardized, creating a gap between clinical and machine learning approaches.
Innovation Solution
The method involves applying constraints to inner nodes of a machine learning network to enforce semantic behaviors, such as correlations with known biomarkers or independence from imaging parameters, allowing for the derivation of data-driven biomarkers that are calibrated and qualified during the training process, using a combination of RADIOME and TECHNOME features to improve image processing and reduce data requirements.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Extent of automation
If deep learning systems are used to automatically derive image features, then feature extraction automation is improved, but interpretability and data efficiency deteriorate
Solution Approach 1:
The patent introduces constrained intermediate representations within the neural network that serve as mediators between automatic feature extraction and interpretability. These constrained layers enforce semantic meanings on hidden representations, allowing the system to automatically learn features while maintaining interpretability through controlled intermediate states that can be linked to known biomarkers or imaging parameters.
2Extent of automation
If deep learning systems are used to automatically derive image features, then feature extraction automation is improved, but data requirements increase
Solution Approach 1:
The patent applies preliminary constraints to the neural network architecture and training process that encode prior knowledge about imaging parameters and biomarker relationships. By pre-defining these constraints before training, the system leverages existing domain knowledge to guide feature learning, reducing the amount of training data needed compared to unconstrained deep learning approaches.
3Loss of information
If hand-crafted features are used in clinical studies, then feature interpretability is improved, but reproducibility and standardization deteriorate
Solution Approach 1:
The patent transforms hand-crafted feature extraction into a learned parameter system where the network automatically optimizes feature representations while constraints maintain interpretability. By changing from fixed hand-crafted parameters to learned parameters with semantic constraints, the system achieves both reproducibility through automated learning and interpretability through enforced semantic meanings on the learned features.
4Measurement precision
If constraints are applied to inner nodes of machine learning network, then biomarker qualification is improved, but system complexity increases
Solution Approach 1:
The patent segments the neural network into distinct functional layers with specific constraints applied to inner nodes. By dividing the network into segments with different constraint types (e.g., some layers constrained to match known biomarkers, others to be independent of imaging parameters), the system achieves precise biomarker qualification while managing complexity through modular constraint application rather than uniform complexity throughout the entire network.
Data Source
AI summary
A method is for obtaining at least one feature of interest, especially a biomarker, from an input image acquired by a medical imaging device. The at least one feature of interest is the output of a respective node of a machine learning network, in particular a deep learning network. The machine learning network processes at least part of the input image as input data. The used machine learning network is trained by machine learning using at least one constraint for the output of at least one inner node of the machine learning network during the machine learning.