Automating DDH Diagnosis using Machine/Deep Learning

Between 2 and 3 percent of all infants are diagnosed with developmental dysplasia of the hip (DDH), and it is believed that approximately 30% of all hip replacement surgeries on patients below the age of 60 are owed to DDH. Graf's Method - the state of the art for diagnosis of DDH in infants - can aid to initiate early on, non-invasive treatment of infants with very high success rates.
Graf's Method relies on measurement of angles between hip bone structures in ultra-sound (US) images. Thus, correct orientation of the ultra-sound probe and the hip of the infant is of fundamental importance for the reliability of the subsequent angle measurements. However, under real-world conditions, acquiring suitable imagery can be difficult and it can as well be a difficult task for an expert to judge the quality of an image when comparably small misalignments are present. In such cases, another examination or the consensus of multiple medical experts might be required to obtain a proper diagnosis.
While previous scientific publications as well as research at SCS has shown that deep learning-based methods can already outperform human experts in measuring angles of the hip bones, there is no method yet to reliably detect automatically whether an US image has been obtained using correct orientation of the US probe.

  • Goals

    We aim to achieve the following goals:

    • Automatically recognizing tilting errors in medical ultrasound images of the hip using Deep Learning (DL) and/or more traditional Machine Learning (ML) algorithms
    • Gathering and preprocessing additional ground truth data from medical experts for real-world US image data
    • Experimental evaluation of whether an automated method is able to surpass the accuracy of an individual human medical expert under certain conditions
  • Tasks

    • Understand the methods in [1, 2, 3] to diagnose DDH
    • Collaborate with medical experts and our team at SCS to gather and preprocess additional ground truth from medical experts on real-life ultra sonic images
    • Either:
      • Design your own method for automated image plane classification using techniques and/or scientific publications of your choice
      • Adapt and retrain the promising methods developed at SCS [2] for automated image plane classification
    • Evaluate and compare the performance of at least one automated method against the performance of human experts
    • Depending on your prior knowledge, preferences, and of course personal interests, special emphasis can be placed on either of:
      • Scientifically-robust study design and evaluation methodology, including literature research and data analysis / visualization
      • Designing and implementing a new approach for detecting titling errors in ultrasound images of the infant hip
      • Shrinking deep neural networks and/or improving the efficiency of the existing approach in [2] (and potentially [1]) to experimentally find an optimal tradeoff point between prediction accuracy and speed for tilting error and DDH detection
    • Document and present the project
      • Written thesis
      • 30 minute presentation


    • [1] Oelen Dave, Implementations of Graf’s Method for DDH Diagnosis Using Deep Learning, Master Thesis at SCS
    • [2] Blatter Mario, Automated Classification of Spatial Orientation of US Images in DDH Diagnosis Using Machine Learning Techniques, Studienarbeit at SCS
    • [3] Golan et al., Fully Automating Graf ’s Method for DDH Diagnosis Using Deep Convolutional Neural Networks, Cham: Springer International Publishing, 2016, pp. 130_141
  • Technologies

    • Machine Learning, Deep Learning, Computer Vision (e.g. scikit-learn, tensorflow, darknet, OpenCV, …)
    • Python and/or C/C++
  • Further Information

    • 60% research, 20% benchmarking, 20% software development
    • Master’s thesis, 1 student
    • Prior knowledge recommended in
      • Good communication skills, ability to pro-actively coordinate yourself with our team and external medical experts
      • Experience in Python and at least basic knowledge of C/C++
    • Prior background in one or more of the following areas will be highly appreciated: data science / statistics, image processing, study design

Haben wir dein Interesse geweckt?

Ich möchte mich auf die Stelle als Automating DDH Diagnosis using Machine/Deep Learning bewerben.