Deep Learning Radiology Report Querying

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Radiology reports are complex and difficult for patients and medical professionals without a medical background to understand, and existing question-answering systems rely on generic information rather than patient-specific data.

Innovation Solution

A machine-learning based system that uses deep learning to train a network to extract patient-specific answers from radiology reports, allowing natural language questions to be answered using evidence from the report, even if the report and questions were not used in training, employing natural language processing and deep machine-learnt networks for parsing and evidence retrieval.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Loss of information

If a comprehensive radiology report is generated with detailed anatomical and medical information, then the diagnostic value and information completeness are improved, but the readability and ease of understanding deteriorate for patients and non-specialist medical professionals

Engineering Contradiction:

Improveinformation completenessVSAvoidreadability

Core Design Contradiction:

Loss of informationVSEase of operation

Solution Approach 1:

The system segments the comprehensive radiology report into structured components (findings, impressions, recommendations) and enables users to query specific segments. The machine-learnt network processes the full report but retrieves only relevant portions to answer user questions, effectively dividing the information retrieval process into targeted segments rather than requiring users to read the entire report.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent introduces an intermediary system (machine-learnt natural language processing network) that mediates between the comprehensive radiology report and the user. This intermediary automatically extracts and presents relevant information in response to user questions, translating the complex medical language into targeted answers without requiring the user to directly interpret the full report.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Ease of operation

If patients use simpler terminology in their questions, then the ease of operation is improved, but the precision of information retrieval may deteriorate

Engineering Contradiction:

Improveease of questioningVSAvoidinformation retrieval precision

Core Design Contradiction:

Ease of operationVSMeasurement precision

Solution Approach 1:

The system applies parameter changes by transforming the question from natural language (potentially with simpler terminology) into a structured query format that the machine-learnt network can process. The network learns to map various terminologies (both simple patient language and complex medical terminology) to the appropriate concepts in the radiology report, maintaining retrieval precision regardless of the question's linguistic complexity.

Inventive Principle:

Principle #35Parameter changes

3Measurement precision

If medical professionals review the entire radiology report to extract specific information, then the accuracy of information extraction is improved, but the time consumption and productivity deteriorate

Engineering Contradiction:

Improveinformation extraction accuracyVSAvoidinformation retrieval efficiency

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The system performs preliminary action by pre-processing and structuring the radiology report during ingestion, organizing the information in a way that enables rapid querying. The machine-learnt network is trained in advance on radiology reports and queries, learning to quickly match user questions to relevant report sections, so that when a professional queries the system, the accurate information is retrieved immediately without requiring manual review of the entire report.

Inventive Principle:

Principle #10Preliminary action

4Adaptability or versatility

If a natural language processing system uses generic information corpus, then the adaptability to different questions is improved, but the accuracy of patient-specific answer extraction deteriorates

Engineering Contradiction:

Improvequestion handling flexibilityVSAvoidpatient-specific answer accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The patent merges two approaches: it uses a machine-learnt network trained on general radiology reports and questions (providing adaptability) while specifically processing each patient's individual radiology report (providing patient-specific accuracy). The system combines the general question-answering capabilities learned during training with the specific patient data at query time, achieving both versatility and precision.

Inventive Principle:

Principle #5Merging (Combining)

Data Source

PatentEP3557584B1Artificial intelligence querying for radiology reports in medical imaging

Publication Date: 2024.08.21 SIEMENS HEALTHINEERS AG

Figure 1
Figure 2
Figure 3~4

AI summary

Information from a radiology report (74) is obtained based on machine learning. Rather than a natural language processing system designed to answer a question based on a large corpus of generic information, deep learning is used to train a machine-learnt network to contribute to extraction (14) of a patient-specific answer from a patient-specific radiology report (74). Natural language questions, including questions with less informed terminology from a patient and questions with more informed terminology from a physician, are answered (20) using evidence from the patient-specific radiology report (74) even where that report and/or questions are not used in training the machine-learnt network.