Last updated: 5/11/18

The focus of this project is to reduce the amount of radiation received by the patient and surgeon during pelvic reconstruction surgery. We enable the use of low-dose images by using a deep learning pipeline to extract structural information from the low-dose image. This concept replaces the need for hundreds of high-dose digital radiographs.

• Students: Mariya Kazachkova, Michael Mudgett
• Mentors: Mathias Unberath, Bastian Bier, Nico Zaech, Nassir Navab, Greg Osgood

Example of results from deep learning pipeline.

In pelvic reconstruction surgery, digital radiographs are used to guide the surgeon through the procedure. This style of imaging provides a clear view of the bones as well as the instrumentation as it is inserted. However, over 100 digital radiogrpahs are typically taken per k-wire inserted and the high dose of radiation inflicted by the x-rays can be dangerous to sensitive areas near the pelvis. Low-dose fluoroscopy is a safer alternative which provides continuous video and is used in many procedures, such as endovascular surgery. Due to the large amount of soft tissue in the abdomen, fluoroscopic images do not give the surgeon a clear view of the workspace.

The goal of our project is to create and train a network which, when fed a low-dose image, can reproduce an analogous high-dose image that a surgeon can make use of in the operating room.

This has two implications:

1. A smaller dose of radiation will be transferred to the patient and surgeon.

2. The imaging process will be sped up and live video will give the surgeon more confidence while placing instrumentation.

• Minimum: (4/1/18)
  1. Simulate a set of x-ray images with varying dose/spectra
  2. Implement and train a de-noising network (code + documentation)

• Expected: (4/22/18)
  1. Functional de-noising pipeline to improve quality of low-dose images (code + documentation)
  2. Chosen dose profile to minimize dose but maximize image quality

• Maximum: (5/6/18)
  1. LSTM network for improving live fluoroscopic video

Our technical approach of this project is to take a large set of simulated x-ray images and use them to train a neural network. The data set will contain images of multiple pelvises at varying angles to make the network robust.

Network:

• Conditional GAN (Generative Adversarial Network) produces estimated high-dose image from input low-dose image
• GAT (Generalized Anscombe Transform) layer stabilizes noise on low-dose image using estimated high-dose image as guide
• Deep CNN learns noise of the stabilized low-dose image
• Inverse-GAT layer returns the denoised low-dose image to its original domain

• MCGPU/PyTorch software – Complete

• GPU access for running MCGPU/training NN – Complete

• CT Volumes for generating images – Complete

1. Milestone name: Set of X-ray Images
   • Planned Date: 3/18/18
   • Expected Date: 3/18/18
   • Status: Complete
2. Milestone name: De-Noising Network
   • Planned Date: 4/1/18
   • Expected Date: 4/1/18
   • Status: Complete
3. Milestone name: Pipeline with VAE
   • Planned Date: 4/22/18
   • Expected Date: 4/22/18
   • Status: Complete
4. Milestone name: LSTM for live fluoroscopy
   • Planned Date: 5/6/18
   • Expected Date: 5/6/18
   • Status: Not being implemented for project

• Project Plan
  • Project Plan Presentation
  • Project Plan Proposal
• Project Background Reading
  • See Bibliography below for links.
• Project Checkpoint
  • Midpoint Presentation
• Paper Seminar Presentations
  • Medical Image Denoising Using Convolutional Denoising Autoencoders -- Presented by Mariya
  • Spektr: A computational tool for x-ray spectral analysis and imaging system optimization -- Presented by Mike
• Teaser Slide
  • Teaser
• Project Final Presentation
  • Final Presentation
• Project Final Report
  • Final Paper

• H. Chen et al., “Low-dose CT denoising with convolutional neural network,” 2017 IEEE 14th International Symposium on Biomedical Imaging (ISBI 2017), Melbourne, VIC, 2017, pp. 143-146.
• J. M. Wolterink, T. Leiner, M. A. Viergever and I. Išgum, “Generative Adversarial Networks for Noise Reduction in Low-Dose CT,” in IEEE Transactions on Medical Imaging, vol. 36, no. 12, pp. 2536-2545, Dec. 2017.
• Dong C., Loy C.C., He K., Tang X. (2014) Learning a Deep Convolutional Network for Image Super-Resolution. In: Fleet D., Pajdla T., Schiele B., Tuytelaars T. (eds) Computer Vision – ECCV 2014. ECCV 2014. Lecture Notes in Computer Science, vol 8692.
• Badal, A. and Badano, A. (2009), Accelerating Monte Carlo simulations of photon transport in a voxelized geometry using a massively parallel graphics processing unit. Med. Phys., 36: 4878–4880.
• A. Badal and A. Badano, “Monte Carlo simulation of X-ray imaging using a graphics processing unit,” 2009 IEEE Nuclear Science Symposium Conference Record (NSS/MIC), Orlando, FL, 2009, pp. 4081-4084.
• J. Baro, J. Sempau, J.M. Fernandez-Varea, F. Salvat, “PENELOPE: An algorithm for Monte Carlo simulation of the penetration and energy loss of electrons and positrons in matter,” Nuclear Instruments and Methods in Physics Research Section B: Beam Interactions with Materials and Atoms, Volume 100, Issue 1, 1995, Pages 31-46.
• L. Gondara, “Medical Image Denoising Using Convolutional Denoising Autoencoders,” 2016 IEEE 16th International Conference on Data Mining Workshops (ICDMW), Barcelona, 2016, pp. 241-246.

• link to code (hosted in private repo on BitBucket): https://bitbucket.org/mkazach1/cis/src/master/