Generating ROC curves for artificial neural networks

Area of Science:

• Medical Imaging
• Machine Learning
• Biostatistics

Background:

• Receiver operating characteristic (ROC) analysis is a standard for evaluating diagnostic performance in medical imaging.
• Artificial neural networks (ANNs) are increasingly used as classifiers in medical imaging, with researchers now reporting ROC curves for them.
• The conventional method for generating ROC curves with ANNs involves adjusting the output node threshold.

Purpose of the Study:

• To propose and evaluate a novel technique for generating ROC curves for two-class ANN classifiers.
• To demonstrate that the proposed method yields superior ROC curves compared to the standard thresholding approach.

Main Methods:

• Development of a new technique for generating ROC curves specifically for two-class ANN classifiers.
• Comparison of the proposed method against the traditional output node thresholding technique.
• Evaluation of generated ROC curves based on area under the curve (AUC) and operating point distribution.

Main Results:

• The proposed technique generates ROC curves with a greater area under the ROC curve (AUC).
• The new method results in a more favorable distribution of operating points on the ROC curve.
• This indicates an improved ability to discern between true positive and false positive rates.

Conclusions:

• The novel ROC curve generation technique offers significant advantages for ANN classifiers in medical imaging.
• This method enhances the assessment of diagnostic performance, leading to more informative ROC analyses.
• The improved AUC and operating point distribution suggest better clinical utility for ANN-based diagnostic tools.