Multimodal Mammographic AI Training for Limited Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing breast cancer analysis systems rely heavily on image data, requiring substantial training datasets that are difficult and costly to acquire due to legal, technical, and workflow barriers, leading to inefficiencies in data anonymization and manual processing.
Innovation Solution
A training procedure and system for artificial intelligence that analyzes multimodal mammographic data, including images, reports, and structured data, using a deep neural network to identify or exclude breast cancer, utilizing cross-learning and pre-training techniques to achieve accurate results with limited data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image-based training data is used for breast cancer analysis, then the system can process visual mammographic information, but substantial training datasets are required which are difficult and costly to acquire
Solution Approach 1:
The patent segments mammographic data into multiple modalities: image data, report texts, and structured data from electronic medical records. This segmentation allows the system to process different types of information separately and combine them, reducing the dependency on large volumes of any single data type while maintaining detection accuracy.
Solution Approach 2:
The system implements a multimodal deep neural network that processes multiple types of input data (images, text reports, structured medical record data) simultaneously. This multi-functional approach allows the system to leverage diverse data sources, reducing the need for extensive single-modality training datasets while improving overall detection performance.
2Measurement precision
If image-based training data is used for breast cancer analysis, then the system can process visual mammographic information, but legal, technical, and workflow barriers make data acquisition costly and time-consuming
Solution Approach 1:
The system processes multiple data modalities including structured data from electronic medical records that can be automatically extracted and anonymized through standard interfaces, reducing manual processing time. Report texts are also processed automatically through natural language processing, further reducing time losses associated with data preparation.
Solution Approach 2:
The system uses text copies of medical reports and structured data extracts from electronic medical records as training inputs. These textual representations can be generated automatically from existing digital records without requiring physical handling or complex anonymization procedures, significantly reducing time losses.
3Ease of manufacture
If manual processing methods are used for mammographic data, then data can be processed with existing tools, but efficiency is reduced due to manual anonymization and tagging
Solution Approach 1:
The patent replaces manual mechanical processing methods with automated computational systems. A deep neural network automatically processes and anonymizes mammographic images, report texts, and structured data. This substitution eliminates manual anonymization and tagging steps, dramatically improving data processing efficiency while maintaining ease of implementation through software-based solutions.
Solution Approach 2:
The system performs self-service data processing where the deep neural network automatically anonymizes images, extracts features from reports and structured data, and prepares training datasets without human intervention. This automation maintains simplicity of implementation while achieving high productivity through the system's autonomous operation.
Data Source
AI summary
Training procedure for artificial intelligence for mammographic data analysis for breast cancer detection including acquiring mammographic data from a plurality of sources and including mammographic images, report texts relating to images, and structured data obtained from SIO, EMR, BI-RADS and MOM including at least metadata relating to part of the images, processing mammographic data through algorithms implementing a multimodal deep neural network (DNN) developing a mammographic data analysis model by performing learning based on sub-phases of first multi-label classification of each image implemented through a model with Encoder-Decoder architecture based on convolutional neural network (CNN) and/or Transformers, association of parts of report texts with images and/or parts of structured data, implemented through a model with Encoder-Decoder architecture based on a bidirectional long-term memory (Bi-LSTM) and/or Transformers, second multi-label classification of mammographic structured data implemented through a model with Encoder-Decoder architecture based on CNN and/or Transformers.