Radiology Report NLP Error Detection via Character-Level Transformations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current methods for assessing the accuracy of radiology diagnoses are subjective, unreliable, and lack a standardized, robust, and quantitative approach, leading to inefficiencies and potential misdiagnoses.

Innovation Solution

A computer-implemented system that utilizes natural language understanding techniques to analyze radiology reports, detect errors, and generate accurate replacements using character-level optical transformation costs and frequency analysis, while also integrating neural networks for named entity recognition and graph convolutions to establish linked relationships.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If natural language understanding techniques with neural networks are used to analyze radiology reports, then diagnostic accuracy assessment becomes standardized and quantitative, but system complexity increases

Engineering Contradiction:
Improvediagnostic accuracy assessmentVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The system segments the radiology report analysis into distinct functional modules: error detection module, character-level optical transformation cost analysis module, frequency analysis module, named entity recognition module, and relationship extraction module. Each module performs a specific task and processes intermediate results, making the overall complex system manageable and maintainable while achieving standardized quantitative assessment

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces intermediary computational structures including character-level embeddings as intermediate representations between input text and final error detection, and attention mechanisms as intermediaries that weigh the importance of different text segments. These intermediaries bridge the gap between raw radiology report text and structured diagnostic accuracy assessments

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If character-level optical transformation costs and frequency analysis are used to detect errors, then error detection precision improves, but computational processing time increases

Engineering Contradiction:
Improveerror detection precisionVSAvoidcomputational processing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary actions by pre-computing character-level embeddings and building frequency analysis data structures before actual error detection. Character-level optical transformation costs are pre-calculated and stored, allowing the main error detection process to query these pre-computed values rather than calculating them from scratch, thus reducing real-time processing time while maintaining high precision

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent applies partial action by focusing computational resources on detecting specific types of errors that are most critical for diagnostic accuracy, rather than attempting to detect all possible errors with equal intensity. The frequency analysis selectively identifies character transformations that are most likely to represent meaningful errors based on their occurrence patterns

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3956901B1Computer-implemented natural language understanding of medical reports
Publication Date: 2025.06.04 COVERA HEALTH
  • EP3956901B1 patent drawingFigure 1
  • EP3956901B1 patent drawingFigure 2
  • EP3956901B1 patent drawingFigure 3

AI summary

A natural language understanding method begins with a radiological report text containing clinical findings. Errors in the text are corrected by analyzing character- level optical transformation costs weighted by a frequency analysis over a corpus corresponding to the report text. For each word within the report text, a word embedding is obtained, character-level embeddings are determined, and the word and character-level embeddings are concatenated to a neural network which generates a plurality of NER tagged spans for the report text. A set of linked relationships are calculated for the NER tagged spans by generating masked text sequences based on the report text and determined pairs of potentially linked NER spans. A dense adjacency matrix is calculated based on attention weights obtained from providing the one or more masked text sequences to a Transformer deep learning network, and graph convolutions are then performed over the calculated dense adjacency matrix.