Genetic Algorithm Feature Selection for Lung Nodule Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current CAD algorithms for detecting lung nodules in MSCT scans face challenges in achieving maximum specificity while maintaining 100% sensitivity, leading to high false positive rates that burden radiologists with unnecessary examinations.
Innovation Solution
The integration of systematic data scaling into genetic algorithm-based feature subset selection systems, which involves choosing a training set, creating a classifier using methods like SVMs or neural networks, and using genetic algorithms to select optimal feature subsets, while removing borderline false positives through Tomek links to reduce false positives and improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If CAD algorithms detect all lung nodules with high sensitivity, then true positive detection is improved, but false positive rate increases requiring more radiologist examinations
Solution Approach 1:
The patent extracts and removes borderline false positive cases from the training set using systematic data scaling. By identifying and excluding these problematic cases, the method creates a cleaner training dataset that reduces false positive rate while maintaining high sensitivity for true nodules.
Solution Approach 2:
The patent implements feedback through iterative training where the classifier performance is continuously evaluated and the training set is dynamically adjusted. This feedback loop allows the system to learn from previous mistakes and progressively improve its ability to distinguish true positives from false positives.
2Reliability
If more training data is used to improve classifier accuracy, then classification performance is improved, but computational time and resource requirements increase
Solution Approach 1:
The patent extracts only the most informative training data by removing borderline false positives and redundant samples. This selective extraction creates a reduced training set that maintains high classification accuracy while significantly reducing computational time requirements compared to using all available data.
Solution Approach 2:
The patent changes the parameter of training set composition through systematic data scaling, transforming the training data from its original form into a scaled and filtered version. This parameter transformation optimizes the balance between training data quantity and computational efficiency.
3Loss of time
If random downscaling is used to reduce training data size, then computational time is reduced, but classification accuracy deteriorates
Solution Approach 1:
The patent applies asymmetric data scaling that treats true positives and false positives differently in the sampling process. Rather than random downscaling, the method systematically scales data based on class balance and case importance, creating an asymmetric transformation that preserves classification accuracy while reducing computational time.
Solution Approach 2:
The patent changes the scaling parameter from random to systematic data scaling, where the scaling factor and selection criteria are dynamically adjusted based on the training data characteristics. This parameter optimization ensures that reduced training sets maintain high classification accuracy.
Data Source
AI summary
Methods and apparatus for training a system for developing a process of data mining, false positive reduction, computer-aided detection, computer-aided diagnosis and artificial intelligence are provided. A method includes choosing a training set from a set of training cases using systematic data scaling and creating a classifier based on the training set using a classification method. The classifier yields fewer false positives. The method is suitable for use with a variety of data mining techniques including support vector machines, neural networks and decision trees.