Skip to content

Q Exchange

Deep learning to characterise ultrasound thyroid nodules

We aim to create an evidence based, deep learning algorithm to characterise thyroid nodules as benign or malignant based on ultrasound appearances and reduce the number of biopsies required.

Read comments 2
  • Proposal
  • 2019

Meet the team

Also:

  • Dr John Canning

What is the challenge your project is going to address and how does it connect to your chosen theme?

Thyroid nodules are a common ultrasound  finding.  The majority are benign.  It is difficult for radiologist to defitively characterise as benign or malignant based on ultrasound images.  Therefore, a large number of biopsies are performed on these patients with the associated risks, discomfort and stress this involves for patients.   In 2018 297 thyroid biopsies were performed at Antrim Area Hospital, 20 of which were malignant.

An evidence based deep learning algorithm will act as

i) an adjunct to the clinical/radiologist decision making tree

ii)reduce the number of biopsies performed

iii)act as a training tool for consultant and trainee radiologists

iv)act as a training tool for ultrasonographers performing head and neck ultrasound.

What does your project aim to achieve?

The aims of the project are to create an evidence based algorithm to characterise thyroid nodules.

We aim to provide an adjunct to the clinical radiological decision making tree.  One potential outcome would be to allow another means of assessing indeterminate (Thy 3a / Thy 3f) nodules to assist thyroid surgeons in deciding wheter or not to operate on these patients.

How will the project be delivered?

We aim to create a dataset of thyroid nodules which have been previously biopsied and characterised as benign / indeterminate / malignant using the cytological Thy1-Thy5 reporting system.

These cases are anonymised.

The ultrasound images are segmented with regions of interest drawn around the nodule which has been biopsied.

A suitable deep learning algorithm will be applied to the anonymised dataset.

I (Dr John Canning) have applied for an MSc Data Analytics at Queen’s University Belfast which commences in September 2019.  I aim to complete the course over 2 years on a part time basis.

https://www.qub.ac.uk/courses/postgraduate-taught/data-analytics-msc/

This course teaches the methods and theory of applying deep learning methods to projects as outlined above.  An ‘industry based project’ represents a considerable component of the course and I aim to complete the above project as part of the MSc.

What and how is your project going to share learning throughout?

This project and the cases contained within the research will act as a comprehensive training set of example cases for consultant and trainee radiologists throughtout the UK and beyond.

Ultrasonagraphers provide an excellent service to patients in our institution, head and neck ultrasound is viewed as a difficult area to gain confidence in, this tool would act as an evidence based adjunct to ultrasonographers beginning to perform and report head and neck ultrasound independently.  Thus helping increase skill mix within radiology departments.

How you can contribute

  • My background is in clinical medicine and radiology, I have a limited computer science / programming background, hence the decision to apply for an MSc in Data Analytics. If there were Q Echange community members who could mentor from a programming (python) or deep learning aspect that would be greatly appreciated.

Plan timeline

30 May 2021 Completion of project and MSc Data Analytics

Comments

  1. Good luck!!!

  2. It would be great to see a more sensitive tool for decision making about the need for a biopsy.  Minimising discomfort for patients and reducing waste of valuable resources in terms of time and cost of test.  Good luck with the bid!

Comments are now closed for this post.