CNN Training With Estimated Region Masks for Tumor Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of region segmentation in machine learning-based image recognition systems, such as pancreatic cancer detection in CT scans, is compromised when using low-quality region data from a secondary recognizer, leading to decreased recognition performance.
Innovation Solution
A training method that combines a first training dataset with ground truth region data and a second dataset with estimated region data to refine the learning model, using a convolutional neural network (CNN) to segment pancreatic tumors in abdominal CT images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If region image data from a secondary recognizer is used as input for training, then the training process can proceed with available data, but the recognition accuracy decreases when the secondary recognizer has low segmentation accuracy
Solution Approach 1:
The patent introduces an intermediary processing step that generates a mask image based on the estimated region image data. This mask serves as a mediator between the low-quality segmented data and the training process, allowing the system to utilize available data while mitigating its poor quality through the masking operation that focuses training on relevant regions.
Solution Approach 2:
The patent transforms the estimated region image data into a different parameter representation by generating a mask image. This parameter transformation allows the training process to work with modified data characteristics that reduce the negative impact of low segmentation accuracy while preserving useful information.
2Quantity of substance
If low-quality region data is used for training, then training can proceed with existing data resources, but the training quality and model performance deteriorate
Solution Approach 1:
The mask image acts as an intermediary that processes the low-quality region data before it is used for training. This intermediary step allows the system to maintain a large quantity of training data while improving its effective quality by filtering and focusing on relevant regions through the mask.
Solution Approach 2:
The patent extracts the essential useful information from the low-quality region data by generating a mask that highlights relevant regions. This extraction process separates the valuable training signals from the noisy or inaccurate portions of the segmented data.
Data Source
AI summary
An information processing apparatus includes at least one memory storing a program, and at least one processor which, by executing the program, causes the information processing apparatus to acquire learning image data, ground truth region image data which indicates a ground truth region of the first region included in the learning image data, and ground truth data on the recognition, acquire estimated region image data which indicates an estimated region of the first region in the learning image data, and train the learning model using a first training dataset constituted of the learning image data, the ground truth region image data and the ground truth data, and a second training dataset constituted of the learning image data, the estimated region image data and the ground truth data.