Transfer learning on ultrasound spectrograms of RF signal for diagnosing nonalcoholic fatty liver disease

Chen, Shi

Permalink

https://hdl.handle.net/2142/104000 Copy

Description

Title

Transfer learning on ultrasound spectrograms of RF signal for diagnosing nonalcoholic fatty liver disease

Author(s)

Chen, Shi

Contributor(s)

• O'Brien Jr., William D.
• Han, Aiguo

Issue Date

2019-05

Keyword(s)

• ultrasound imaging
• radiofrequency signal
• spectrograms
• deep learning
• transfer learning
• NAFLD

Abstract

Purpose: Nonalcoholic fatty liver disease (NAFLD) is a very common liver disease and affects 20-30% of the world population. The gold standard for diagnosing NAFLD is the liver biopsy. The purpose of this thesis is to develop and evaluate a 2D convolutional neural network (CNN) which uses the ultrasound spectrograms of radiofrequency (RF) signals for diagnosing NAFLD. Spectrogram of an RF signal shows the frequency spectrum of the RF signal as it varies with time. Methods: 204 research participants with and without NAFLD took both ultrasound imaging and MRI. The results given by the MRI-proton density fat fraction (PDFF) were used as the reference (NALFD defined as PDFF≥5%). For ultrasound images from all patients, a fixed region-of-interest (ROI) was chosen to maximize the portion of the liver in the ROI. The time-gain-compensation (TGC) was removed from the RF signals and RF signals were downsampled and normalized. For each RF signal, a spectrogram with size 32×32 was generated and the log-transformed absolute values of spectrograms were used as inputs for the 2-D CNN model. A transfer learning approach was used, where a pretrained VGG-19 model was used as the base model and was fine-tuned by training the last few convolution layers and fully connected layers to diagnose NAFLD. The hyperparameter optimization was done on the training set by a grid search to optimize the number of trainable convolution layers and the number of nodes in the first fully connected layer. Results: The model achieved a sensitivity of 94.3%, a specificity of 90.6% and an accuracy of 93.1%s for diagnosing NAFLD on the test set. Conclusions: It is feasible to develop a 2D model from transfer learning on ultrasound spectrograms of RF signal to accurately diagnose NAFLD. The proposed approach provides a noninvasive method to evaluate the fatty liver disease.

Type of Resource

text

Language

en

Permalink

http://hdl.handle.net/2142/104000

Owning Collections