Multi-task Adapters for Pathology Extraction from Radiological Reports
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for diagnosing pathologies from medical reports are inefficient due to variations in writing styles and lack of standardized data, leading to high computational and deployment costs, and errors in diagnostic accuracy.
Innovation Solution
Implementing a multi-task learning approach using a Bidirectional Encoder Representations from Transformers (BERT)-based machine learning network with multi-task adapters to generate pathology-specific models for efficient extraction of pathology and severity information from radiological reports, reducing the need for separate models for each pathology and body part.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate machine learning models are trained for each pathology and body part, then diagnostic accuracy is improved, but computational cost and deployment complexity increase significantly
Solution Approach 1:
The patent combines multiple separate pathology detection models into a single unified multi-task learning model that can simultaneously detect multiple pathologies across different body parts. This consolidation maintains diagnostic accuracy while significantly reducing deployment complexity and computational overhead by sharing common features and representations across tasks.
Solution Approach 2:
The patent creates a universal base model that serves multiple functions by implementing multi-task learning capabilities. The single model can adapt to detect various pathologies in different anatomical regions through task-specific adapters, eliminating the need for separate specialized models and enabling one system to perform multiple diagnostic functions.
2Reliability
If traditional separate models are used for each pathology, then task specificity is achieved, but training time and computational resources are excessively consumed
Solution Approach 1:
The patent implements a two-stage training approach where a general base model is pre-trained on diverse medical imaging data first, establishing foundational features and representations. Then, task-specific adapters are fine-tuned on individual pathology datasets. This preliminary action reduces overall training time compared to training separate models from scratch while maintaining task specificity.
Solution Approach 2:
The patent segments the model architecture into a shared base model and task-specific adapter components. The base model learns general medical imaging features that transfer across tasks, while adapters capture task-specific nuances. This segmentation allows efficient knowledge transfer and reduces redundant training while preserving task specificity.
3Measurement precision
If extensive labeled data is collected for each pathology, then model performance is improved, but data collection and annotation costs increase
Solution Approach 1:
The patent changes the learning parameters by implementing multi-task learning with shared representations. This allows the model to learn from diverse pathology data simultaneously, improving performance on individual tasks with fewer labeled examples per pathology by leveraging patterns learned from related tasks and shared visual features.
Data Source
AI summary
Described are techniques for processing text data. A process can include obtaining an input text data indicative of pathologies associated with a corresponding radiological image. A BERT-based machine learning network can be used to generate a plurality of location tags, each location tag associated with a sentence of input text data and indicative of an anatomical location in the corresponding radiological image. A plurality of sentence groups can be generated using the input text data and the plurality of location tags, each sentence group including sentences of input text data that are associated with a location tag indicative of the same anatomical location. A multi-task learning (MTL)-based machine learning network can be used to generate a plurality of sets of features, each set of features generated based on a particular sentence group and indicative of pathology severity predictions determined for the anatomical location associated with the particular sentence group.