- Title Information
- Title
- Quantitative computational estimation of liver steatosis based on parameters
derived from diagnostic ultrasound
- Name:
Personal
- Name Part
- Tuomi, Adam
- Role
- Role Term:
Text
- author
- Name:
Personal
- Name Part
- Wu, Jie Ying
- Role
- Role Term:
Text
- author
- Name:
Personal
- Name Part
- Beland, Michael
- Role
- Role Term:
Text
- author
- Name:
Personal
- Name Part
- Konrad, Joseph
- Role
- Role Term:
Text
- author
- Name:
Personal
- Name Part
- Glidden, David
- Role
- Role Term:
Text
- author
- Name:
Personal
- Name Part
- Grand, David
- Role
- Role Term:
Text
- author
- Name:
Personal
- Name Part
- Merck, Derek
- Role
- Role Term:
Text
- advisor
- Name:
Corporate
- Name Part
- Brown University.
- Name Part
- Alpert Medical School
- Name Part
- Scholarly Concentration Program
- Name Part
- Medical Technology and Innovation
- Role
- Role Term:
Text
- affiliation
- Origin Information
- Date Created
(keyDate="yes", encoding="w3cdtf")
- 2015-11
- Note
(displayLabel="Scholarly concentration")
- Medical Technology and Innovation
- Abstract
- In current medical practice, an ultrasound guided, random needle biopsy of the
liver is the gold standard in hepatic steatosis assessment. Ultrasound imaging is used to
place a needle into the patient’s liver, and extract a core of tissue for pathologist
examination. This procedure carries a risk of potentially life threatening bleeding, and thus
substitution of a needle biopsy with a non-invasive alternative could reduce adverse events.
We have developed a machine learning algorithm for analyzing ultrasound (US) images
quantitatively to provide computer-aided diagnosis of hepatic steatosis. We built the
algorithm using liver US studies from 288 patients, and correlated to their corresponding
biopsy assessments. Radiologists identified a region of interest (ROI) on each image which was
then filtered for various texture responses. These texture responses formed the
parameterization for the machine learning algorithms which, along with the pathology-confirmed
diagnoses, were used to train classifiers. Testing with cross-validation, we were able to
classify US images as steatotic or normal with sensitivities of 40-74%, and specificities of
72-86%
- Subject (Local)
- Topic
- Machine learning
- Subject (Local)
- Topic
- Diagnostic imaging
- Subject (Local)
- Topic
- Computer-aided diagnosis
- Type of Resource
- text