Deep Learning Model Training via Co-Registered Histopathological Image Labelling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge in developing deep learning models for histopathological image analysis is the time-consuming and subjective nature of manual ground truth labelling, which is prone to variability, especially when using less accurate and cheaper staining methods like H&E, limiting the effectiveness of these models when they rely on more expensive IHC-stained sections.
Innovation Solution
A method that uses co-registration of adjacent tissue sections stained with different protocols, such as IHC and H&E, to automatically transfer histopathological information, allowing a deep learning model to recognize specific IHC-based staining information in H&E-stained samples, thereby reducing the need for expensive staining and minimizing human subjectivity in labelling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual ground truth labelling is used by human experts, then histopathological information can be obtained, but the process is time-consuming and prone to human subjectivity and inter-annotator variability
Solution Approach 1:
The system uses automatically generated ground truth labels from the same deep learning model being trained, creating a self-service labelling loop. The model generates initial labels, these labels are used to train the model further, eliminating the need for manual expert labelling while maintaining consistent labelling standards throughout the process.
2Measurement precision
If expensive IHC staining methods are used to obtain accurate histopathological information, then labelling accuracy improves, but the cost increases and the method cannot always be used
Solution Approach 1:
The system creates virtual copies of IHC staining information by using the deep learning model to generate ground truth labels that replicate the histopathological information normally obtained from expensive IHC staining. These virtual labels allow the model to learn from IHC-equivalent data without requiring actual IHC-stained sections.
Solution Approach 2:
The system changes the staining parameter from expensive IHC methods to cheaper H&E staining, while compensating for the reduced information quality through the use of automatically generated ground truth labels. This allows the model to be trained on abundant, inexpensive H&E images while maintaining the accuracy equivalent to IHC-based labelling.
3Ease of manufacture
If cheaper H&E staining methods are used, then cost decreases and availability increases, but the staining provides less histopathological information and is less accurate
Solution Approach 1:
The deep learning model acts as an intermediary that translates information from H&E stained images into accurate histopathological labels. The model learns to recognize patterns in H&E images and generates ground truth labels that compensate for the inherently less information-rich H&E staining, effectively bridging the gap between cheap staining and accurate information extraction.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
The present disclosure relates to a method for training a deep learning model to obtain histopathological information from images, the method comprising the steps of: providing a first image of a first section of a specimen, such as a tissue portion, wherein the first section has been stained using a first staining protocol; providing a second image of a second section of the specimen, the first and second sections being adjacent sections of the specimen, wherein the second section has been stained using a second staining protocol different from the first staining protocol; co-registration of the first and second sections of the images; obtaining histopathological information for the first section of the first image based on the staining of the first section and the first staining protocol; transferring the histopathological information for the first section of the first image to the second section of the second image based on the co- registration of the first and second sections, thereby labelling the second image according to the histopathological information of the first image; and training a deep learning model, based on the labelling of the second image, to obtain histopathological information from images stained with the second staining protocol. The disclosure further relates to a system for training a deep learning model, comprising a computer- readable storage device for storing instructions that, when executed by a processor, performs the method for training a deep learning model to obtain histopathological information from images.