Breast lesion classification using ultrasound and elastrographic images

Abstract

Machine Learning has shown good performance in large-scale medical-related datasets with thousands of records. However, adopting machine learning in medical domains with smaller datasets raises concerns. One of them is caused by the uncertain performance in such datasets. In this study, analyses based on 187 patients were performed to examine the feasibility of classifying breast lesion with machine learning models in practice. The study contained three stages, including feature extraction, dimensionality reduction and model comparison. Features were extracted from the ultrasound and elastography images based on their Region of Interest before different dimensionality reduction methods, such as PCA, Isomap, Variance Threshold and Chi-squared tests, were performed. Several classic machines learning models, including SVM, Random Forest, Stochastic Gradient Descent and more, were trained and evaluated on the preprocessed data. After evaluation, single dimensionality reduction methods result in good performances. However, by combining different dimensionality reduction methods, a satisfying result with an accuracy of 0.967 can be achieved. This research provides a thorough study in breast lesion classification based on a single dataset. With the methodology and results obtained in this research, the study supports that machine learning could assist in real-life medical application.