UCL Centre for Medical Image Computing


Adam Szmul and Kyriaki-Alkisti Stavropoulou - CMIC/WEISS Joint Seminar Series

10 November 2021, 1:00 pm–2:00 pm

Adam Szmul and Kyriaki-Alkisti Stavropoulou - talks as part of CMIC/WEISS Joint Seminar Series

Event Information

Open to





UCL Centre for Medical Image Computing and Wellcome/EPSRC Centre for Interventional and Surgical Sciences

Speaker: Adam Szmul

Title: From domain adaptation to radiation induced lung damage tissue classification, applications of ML in radiotherapy


Machine learning methods become increasingly widely used in radiotherapy (RT). In our work we applied them in two tasks: to perform domain adaptation from Cone Beam Computed Tomography (CT) to CT, and in developing and automating lung tissue classification system for Radiation-Induced Lung Damage (RILD).

The Cone Beam Computed Tomography (CT) to CT synthesis framework is tailored for paediatric abdominal RT patients. Our approach was based on the cycle-consistent adversarial generative network (cycleGAN) modified to preserve structural consistency. To adjust for differences in field-of-view and body size from different patient groups, our training data was spatially co-registered to a common field-of-view and normalised to a fixed size. In the proposed synthesis pipeline only global residuals are learned and predicted. This approach allows one to refine the raw CBCTs by removing the unwanted artefacts, rather than generating new images “inspired” by the input. The proposed framework showed improvements in generating synthetic CTs from CBCTs compared to the original implementation of cycleGAN

Radiation-Induced Lung Damage (RILD) is a common side effect of RT in lung cancer patients. The ability to automatically segment, classify and quantify different types of lung parenchymal change is essential to uncover underlying patterns of RILD and their evolution over time. A RILD dedicated tissue classification system was developed to describe lung parenchymal tissue changes on a voxel-wise level. The classification system was automated for segmentation of five lung tissue classes on CT scans which described incrementally increasing tissue density, ranging from normal lung to consolidation. For ground truth data generation, we initiated a two-stage data annotation approach, akin to active learning. The final auto-segmentation algorithm was an ensemble of six 2D-Unets using different loss functions and numbers of input channels. We performed quantitative and qualitative evaluation of our approach.


Speaker: Kyriaki-Alkisti Stavropoulou

Title: A multichannel feature-based approach for longitudinal lung CT registration in the presence of radiation induced lung damage


Quantifying parenchymal tissue changes in the lungs is imperative in furthering the study of radiation- induced lung damage (RILD). Registering lung images from different time-points is a key step of this process. Traditional intensity-based registration approaches fail this task due to the considerable anatomical changes that occur between timepoints. This work proposes a novel method to successfully register longitudinal pre- and post-radiotherapy (RT) lung CT scans that exhibit large changes due to RILD, by extracting consistent anatomical features from CT (lung boundaries, main airways, vessels) and using these features to optimise the registrations.



Chair: Jamie McClelland