In Current medical imaging
INTRODUCTION : This study demonstrates the possibility of detecting tumors on mammograms with high accuracy (more than 72%) using neural networks and studies the characteristics of machine learning models for improving their efficiency.
METHOD : We have proposed image preprocessing methods that enable high classification accuracy, as well as methods of increasing the training set and balancing the distribution of diagnostic classes when the training set is small. The classification has been done for the following four diagnostic classes: dysplasia, pre-cancer state (ductal carcinoma in situ), cancer state (invasive carcinoma), and benign tumor.
RESULTS AND CONCLUSION : We have conducted experiments to compare different models based on convolution neural networks and proposed methods for estimating the model quality. We have obtained a base model that can be used to make recommendations to establish a diagnosis. We have studied the characteristics of the base model and identified promising directions of modification for further improving the quality estimates.
Ivanova Galina S, Golovkov Alexander A, Petrova Iana S, Borodin Alexander A, Shakhlan Anastasia O, Umnov Alexander V, Lonshakova Kristina A, Kelenin Vladimir V
Cancer diagnostics, deep machine learning, mammogram, neural networks