Deep CNN Patch Decomposition for Abnormality Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for detecting abnormalities in anatomical images, such as cancer, are time-consuming and prone to errors due to the high variability of tissue patterns and the need for exhaustive manual annotations, which are costly and difficult to obtain.
Innovation Solution
A deep convolutional neural network (CNN) is trained using weakly labeled anatomical images, where images are decomposed into patches, and the network computes a probability of abnormality based on the maximal probability value of each patch, allowing for automatic detection and localization of abnormalities without requiring explicit annotation of the abnormality's location.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If manual analysis of anatomical images is performed by radiologists, then detection accuracy can be maintained through expert judgment, but the process becomes time-consuming and tedious
Solution Approach 1:
The patent replaces the mechanical system of manual radiologist analysis with an automated deep learning system. The CNN processes anatomical images automatically, substituting human expert judgment with machine-based pattern recognition. This substitution maintains detection accuracy through trained neural networks while dramatically reducing processing time and eliminating manual labor.
2Measurement precision
If exhaustive manual annotations are used for training, then model accuracy improves, but the cost and difficulty of obtaining annotations increases significantly
Solution Approach 1:
The patent applies segmentation by dividing the training process into two stages: first training on weakly labeled data (image-level annotations only), then refining with selectively applied strong labels. This segmentation allows the model to achieve high accuracy without requiring exhaustive manual annotations for all training images, significantly reducing annotation costs while maintaining model precision.
Solution Approach 2:
The patent uses partial action by applying exhaustive annotations only to a subset of training images (those identified as having abnormalities) rather than all images. This partial annotation strategy, combined with weakly supervised learning on the full dataset, achieves high model accuracy while dramatically reducing the total annotation effort and cost required.
3Measurement precision
If high-resolution images are processed, then detection accuracy improves, but computational resources and processing time increase
Solution Approach 1:
The patent segments high-resolution anatomical images into multiple smaller patches before processing. This segmentation allows the CNN to analyze fine-grained details in each patch at full resolution, maintaining detection accuracy for small abnormalities while reducing the computational burden of processing entire high-resolution images as single units. The patch-based approach distributes computational resources more efficiently.
Data Source
AI summary
There is provided a method for training a deep convolutional neural network (CNN) for detecting an indication of likelihood of abnormality, comprising: receiving anatomical training images, each including an associated annotation indicative of abnormality for the whole image without an indication of location of the abnormality, executing, for each anatomical training image: decomposing the anatomical training image into patches, computing a feature representation of each patch, computing for each patch, according to the feature representation of the patch, a probability that the patch includes an indication of abnormality, setting a probability indicative of likelihood of abnormality in the anatomical image according to the maximal probability value computed for one patch, and training a deep CNN for detecting an indication of likelihood of abnormality in a target anatomical image according to the patches of the anatomical training images, the one patch, and the probability set for each respective anatomical training image.