Medical Image Tiling With Ranking Loss for Cancer Grading
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning methods for cancer grading in medical imaging fail to capture relative severity relationships, which are crucial for tasks like cancer diagnosis and treatment planning, particularly in prostate cancer grading where Gleason scores provide valuable information about cancer severity.
Innovation Solution
A system that processes electronic medical images by dividing them into tiles, predicting cancer sub-category proportions using a machine learning model trained with ranking loss, and determining an overall grade by formulating 'greater than' and 'less than' relations between Gleason patterns, converting these into mathematical equations to train the model effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If traditional machine learning classification or regression methods are used for cancer grading, then the system is simple to implement, but it fails to capture relative severity relationships between different cancer patterns
Solution Approach 1:
The patent changes the learning objective parameter from standard classification loss or regression loss to ranking loss. This parameter change enables the model to capture relative severity relationships by learning to rank different cancer patterns according to their severity, rather than simply classifying them into discrete categories or predicting continuous values without ordinal relationships.
Solution Approach 2:
The patent introduces ranking as an intermediary concept between traditional classification and the goal of capturing relative severity. By formulating the problem as a ranking task where the model learns to order cancer patterns by severity, the system captures relational information that neither simple classification nor regression alone can provide.
2Measurement precision
If the system divides images into tiles and predicts proportions of cancer sub-categories, then it can capture relative severity relationships, but the computational complexity increases
Solution Approach 1:
The patent segments the medical image into multiple tiles, allowing the model to process local regions independently and predict cancer pattern proportions for each tile. This segmentation enables precise local characterization of cancer severity while the overall grade is determined by aggregating tile-level predictions, balancing accuracy with computational feasibility.
Solution Approach 2:
Instead of analyzing the entire image globally, the system performs partial analysis on individual tiles. This partial action approach reduces computational power requirements compared to global analysis, while still achieving high measurement precision through aggregation of local predictions.
3Loss of information
If the machine learning model is trained with ranking loss to capture Gleason pattern relationships, then the cancer grading becomes more informative, but the training process becomes more complex
Solution Approach 1:
The patent changes the training objective parameter from standard loss functions to ranking loss. This parameter change enables the model to learn proportion relationships between Gleason patterns by optimizing rank-based objectives, capturing the relative severity information encoded in Gleason scoring systems while providing a clear training signal for the optimization process.
Data Source
AI summary
A computer-implemented method for processing medical images, the method comprising receiving a plurality of medical images of at least one pathology specimen, the pathology specimen being associated with a patient. The method may further comprise dividing the one or more medical images into a plurality of tiles and predicting, using a machine learning system, proportions of each type of cancer sub-category for the plurality of tiles, the machine learning system having been trained by ranking loss. The method may further include determining an overall grade of cancer for the one or more medical images.