Classification of lesions in breast ultrasound images using neural networks

View/Open

In this project we aimed to evaluate and build a neural network framework for classification of lesions in breast ultrasound images. We built a classification model based on k-Nearest Neighbour ((kNN) classifier to serve as an evaluation baseline. We then built 4 neural network models using the TensorFlow and keras deep learning libraries; A fully connected neural network, a custom convolutional neural network (CNN) and two transfer learning networks based on retraining InceptionV3 which is a state of the art general purpose image classification CNN. Neural network approaches outperformed the kNN. Our best CNN manages to achieve a low false negative rate and high true positive rate hence a high sensitivity of 0.85. The transfer learning networks underperform due to data limitations. We recommend more data be used in future studies to fully investigate the potential of transfer learning for ultrasound lesion classification. (2.807Mb)

Author

OGWAL, Emmanuel Hian

xmlui.mirage2.itemSummaryView.MetaData

Show full item record

Please cite this item using this URL

http://hdl.handle.net/20.500.12281/4226

xmlui.mirage2.itemSummaryView.Collections

Academic submissions (CEDAT)