Machine learning scatter estimation for dental cone-beam CT


Internship objective

The goal of this internship is to design a  model capable of estimating CT scatter in dental cone-beam CT imaging from the acquired projections only. The candidate will compare an interpretable Gaussian convolution model with a deep learning approach, evaluate their robustness in the presence of metal, and determine whether improved scatter simulation benefits downstream medical image processing tasks.

Scientific Context

The development of supervised learning methods for medical image processing depends on the availability of sufficiently large datasets associated with reference ground truth. In medical imaging, obtaining such data is particularly challenging because of cost and practical limitations. Physics-based simulation provides an alternative by generating controlled data under different acquisition and material conditions.

This internship focuses on X-ray computed tomography (CT). During a CT acquisition, X-ray photons are emitted by a source, pass through the patient, and are measured by a detector. The measured signal depends primarily on the attenuation of the X-rays by the materials encountered along their paths. Ideally, the detector would measure only primary photons, meaning photons that travel directly from the source to the detector.

In practice, some photons interact with matter and change direction before reaching the detector. These photons contribute an additional signal known as scatter. Scatter modifies the measured projections and can lead to biased intensity values. Consequently, estimating scatter is an important component of realistic CT simulation.

Monte Carlo methods can simulate scatter by modeling a large number of individual photon trajectories and stochastic interactions with matter. These methods provide physically realistic estimates and can therefore be used to generate reference scatter data. However, their high computational cost makes them unsuitable for producing scatter estimates for every sample in a large training dataset.

A faster alternative is to learn a trainable model from a limited number of Monte Carlo simulations. Gaussian convolution models provide an efficient and interpretable approximation, while deep learning models may capture more complex relationships at the cost of increased model complexity and training requirements.

A key challenge is the presence of high-attenuation materials such as metal, which can produce data that differ  from standard cases. This raises an important generalization question: can a model trained on metal-free data accurately estimate scatter in cases containing metal?

The objective of this internship is to compare Gaussian convolution and deep learning models for approximating Monte Carlo scatter simulations. The study will evaluate their accuracy, generalization to metal-containing data, computational efficiency, and potential impact on downstream image-processing tasks.

 


Internship Work Plan 

The intern will join the Tomoradio team at CREATIS and work on the following tasks:

  • Literature review Review existing algorithms for scatter estimation in X-ray CT, including Gaussian convolution and deep learning approaches. 
       
  • Data preparation Analyze and preprocess the CT projection data and Monte Carlo scatter references, considering cases with and without metal. 
       
  • Model development Implement and train a Gaussian convolution model and a deep learning model for scatter estimation. 
       
  • Generalization and performance evaluation Compare both models with Monte Carlo references and assess their ability to generalize to high-attenuation regions containing metal. 
       
  • Downstream evaluation Determine whether improved scatter estimation benefits a supervised image-processing task, such as tooth segmentation on cone-beam CT images.
       
     

Technologies and Expected Skills

  • Python   
  • NumPy and SciPy   
  • PyTorch   
  • Scikit-learn
  • ITK and RTK would be an advantage but are not required.
     

A background in computer sciences, applied mathematics, physics, or image processing is welcome. Knowledge of medical imaging or CT physics is welcome but not mandatory. The candidate should be comfortable with Python programming, numerical experimentation, data analysis, and the evaluation of machine learning models.

References

  • B Ohnesorge, T Flohr, and K Klingenbeck-Regn. Efficient object scatter correction algorithm for third and fourth generation CT scanners. European radiology, 9(3):563–569, 1999
  • Lukas Hennemann, Julien Erath, Andreas Heinkele, Eric Fourni´e, Martin Petersilka, Karl Stierstorfer, and Marc Kachelrieß. Spectral deep learning-based patient and bowtie scatter correction for clinical photon-counting ct. Medical Physics, 53(5):e70442, 2026.
  • Miguel Sun and Josh Star-Lack. Improved scatter correction using adaptive scatter kernel superposition. Physics in medicine and biology, 55:6695–720, 10 2010
  • Heesin Lee and Joonwhoan Lee. A deep learning-based scatter correction of simulated x-ray images. Electronics, 8(9):944, 2019