Learning Angles from Ultrasonography Imagery with Machine Learning and Deep Learning

Pediatricians and general practitioners use ultrasound for diagnostic purposes in their office. Ultrasound is fast, reliable, safe, and is seen as the visual stethoscope of the 21st century. Standardized, high-quality examinations depend on the exact position of the transducer for transmitting and receiving the ultrasound beam. However, the correct plane defining the image cross section obtained from the patient is often difficult to achieve.Advanced technology might support physicians with feedback based on automated analyses of examinations by ultrasound. We will use the example of ultrasound-based diagnosis of developmental dysplasia of the hip (DDH) in newborns to suggest such an approach using Neural Networks.

Goals
We aim at achieving the following goals:

  • Automatically recognizing tilting errors in medical ultrasound images of the hip using Deep Learning (DL) and Machine Learning (ML) algorithms
  • Using state-of-the-art deep neural network architectures (CNN, Inception, ResNet, etc.) to improve the predictive performance of the current DDH diagnostic algorithms (beyond human accuracy)
  • Reducing the size of the deep neural network to find an optimal tradeoff between prediction accuracy and speed
  • Integrating the DL/ML algorithm into a software prototype that enables real-time detection of tilting errors

Tasks

  • Understand the method in [1, 2] to diagnose DDH
  • Reproduce the results in [1] for automated DDH detection
  • Implement as algorithm prototypes
    • DL and ML algorithms for detecting titling errors in ultrasound images of the hip
    • State-of-the-art DL algorithms to supersede the current state-of-the-art DDH algorithm to achieve superhuman accuracy
    • Shrink deep neural networks to find optimal tradeoff point between prediction accuracy and speed for tilting error and DDH detection
  • Implement and test as software prototype
    • Implement the most promising DL/ML algorithm in a software prototype for real-time detection of tilting errors and DDH diagnosis
  • Document and present Master Thesis
    • Final document
    • One-Minute Video Summary
    • 30 minute presentation

Technologies

  • Machine Learning, Deep Learning, Computer Vision
  • Python, TensorFlow, scikit-learn
  • Optional: R, C++, MATLAB, MxNet

References

  • [1] Oelen D, Implementations of Graf’s Method for DDH Diagnosis Using Deep Learning, Master Thesis at SCS
  • [2] Golan et al., Fully Automating Graf ’s Method for DDH Diagnosis Using Deep Convolutional Neural Networks, Cham: Springer International Publishing, 2016, pp. 130_141

Kind of Work
60% Research, 20% Benchmarking, 20% Software Development

Requirements
Machine Learning + Deep Learning Basics, Python, Matlab

Time & Effort
Master’s Thesis, 1 Person

Contact
fabian.schenkel@scs.ch

SCS - Studienarbeiten

Supercomputing Systems AG  •  Technoparkstrasse 1  •  CH-8005 Zürich  •  Phone +41 43 456 16 00  •  Fax +41 43 456 16 10  •  info@scs.ch  •  Impressum  •  Datenschutzerklärung