Pathological Image Caption Matching for Automatic Subtype Diagnosis
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional technologies for diagnosing lesions such as dermatitis and cancer lack accuracy in determining subtypes based on pathological images, relying heavily on a doctor's experience and skills, and do not provide sufficient automated assistance.
Innovation Solution
A diagnosis assistance apparatus and method utilizing machine learning to generate captions for pathological images in a predetermined format, evaluate similarities with database findings, and output the most similar subtype, enhancing accuracy through structured comparisons.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional image retrieval technology is used to support diagnosis, then similar images can be extracted from the database, but the accuracy in deriving features is insufficient and doctors still need to refer to atlases
Solution Approach 1:
The patent replaces the manual mechanical process of doctors referring to atlases and visually comparing images with an automated computer vision system. The system uses deep learning models to automatically extract features, generate captions, and determine subtypes, substituting the mechanical manual review process with automated image processing and analysis.
Solution Approach 2:
The system enables self-service diagnosis assistance by automatically processing pathological images through multiple learned models. The apparatus autonomously performs feature extraction, caption generation, similarity comparison, and subtype determination without requiring manual intervention or reference to external atlases, making the diagnostic support system self-sufficient.
2Reliability
If doctors use conventional techniques with atlases to determine subtypes, then determination can be made, but it relies on doctor's experience and skills resulting in problems in accuracy and efficiency
Solution Approach 1:
The patent replaces the reliance on doctors' subjective experience and skills with an automated system based on deep learning models. The system objectively analyzes pathological images through trained neural networks, eliminating variability in human expertise and providing consistent, reproducible subtype determinations that do not depend on individual doctor qualifications.
Solution Approach 2:
The system performs preliminary analysis by pre-processing images through multiple learned models that extract features and generate captions before final subtype determination. This preliminary processing prepares the data in advance, enabling faster and more accurate comparison against the database, thereby reducing overall diagnosis time while improving reliability.
3Extent of automation
If manual reference to atlases is required for final determination, then subtype can be determined, but the process is inefficient and lacks automation
Solution Approach 1:
The system achieves full automation by making the diagnosis assistance apparatus self-sufficient. It automatically extracts features from pathological images, generates descriptive captions, compares results with the database, and determines subtypes without requiring manual reference to atlases or human intervention at any stage, thereby maximizing both automation extent and diagnostic efficiency.
Solution Approach 2:
The patent merges multiple previously separate diagnostic steps into a single integrated automated system. Feature extraction, caption generation, similarity comparison, and subtype determination are combined into one unified process that operates automatically, eliminating the need for separate manual operations and significantly improving diagnosis efficiency.
Data Source
AI summary
In order to attain an object to improve accuracy in automatic determination of a subtype carried out on the basis of a pathological image, at least one processor included in a diagnosis assistance apparatus carries out: a first acquisition process of acquiring a first caption from a first learned model, the first learned model being constructed by machine learning so as to generate, in a case where a pathological image is inputted, a caption describing content of the pathological image in a predetermined format, the first caption describing, in the predetermined format, content of a first pathological image to be subjected to diagnosis; a first evaluation process of evaluating a first similarity which is a similarity between (i) content of at least a part of a plurality of findings which are accumulated in a database and which represent, in writing in the predetermined format, a respective plurality of pathological subtypes and (ii) content of the first caption; and an output process of outputting information pertaining to a subtype whose first similarity is evaluated to be the highest among those of the plurality of pathological subtypes. The information outputted in the output process is used for decision making in diagnosis by a doctor.