Alexander Leemans

UMC Utrecht

In short, Alexander Leemans' research entails the following: 

Harmonization and segmentation of multicenter paediatric sarcoma MRI data through artificial intelligence to support clinical decision making in international trials

'Rhabdomyosarcoma and Ewing sarcoma are rare tumours with a prevalence of around 35 cases per year in the Dutch paediatric and adolescent population. Currently, there are no biomarkers that can distinguish between poor and good outcomes at an early stage of treatment of these tumours. While microstructural tissue properties derived from diffusion MRI (dMRI) could be valuable biomarkers for early assessment of treatment efficacy, differences in scanner hardware and acquisition protocols across sites complicate dMRI data pooling necessary for clinical decision making in international trials. In this project, we will develop artificial intelligence (AI) methodology for harmonizing dMRI signals and segmenting tumour MRI data that can alleviate unwanted multicenter variability while preserving the biological signature of tissue microstructure.'

 

Publications

• Evaluation of harmonization strategies for analyzing multi-center diffusion MRI data of rhabdomyosarcomaISMRN & ISMRT Annual Meeting & Exhibition, Mei 2025
• Klik hier om de publicatie te bekijken

 

Brief summary of progress / results October 2025

Extensive experiments were performed with changing hyperparameters of the image segmentation model. The modifications to rule-based parameters yielded significant improvements, while most augmentation strategies decreased segmentation accuracy. The results have been described in a paper that is being finalized for submission to a radiology journal.

Additionally, we experimented with employing DinoV2 as an encoder part of the segmentation model for the head-and-neck patients. The model achieved a Dice score of ~0.6 when validating on subjects from centres that were not present in the training dataset.

The group has been invited to contribute to “Prompting the future: artificial intelligence in paediatric radiology” collection for Paediatric Radiology, with a proposed paper title: “Role of artificial intelligence in paediatric oncologic imaging”. The paper is currently in preparation for submission. 

The PhD student obtained an additional dataset from Princess Máxima Centre. The dataset is expected to improve the heterogeneity of the EpSSG RMS 2005 study dataset, which has been used so far. Currently, curation and segmentation of the data are being performed.

Regarding the harmonization project, we compared the performance of ComBat harmonisation algorithm with the harmonisation based on General Linear Model regression. The effect of harmonisation on the correlation with patient survival was investigated. The results were presented at the Annual Congress of the International Society for Magnetic Resonance in Medicine. Experiments were mostly focused on the headand-neck patients.

We repeated the data curation pipeline of van Ewijk, Chatziantoniou et al to observe whether harmonisation changes the relationships between distributions of alveolar and embryonal rhabdomyosarcoma cases. We also added quantile normalisation to the set of harmonisation algorithms. Statistical tests have not shown any changes. However, when exploring the effect of harmonisation on the connection of MRI-based features with survival, the results appear to be unreliable, given the inappropriate size effect. We will compensate for that challenge by using the dataset from Princess Máxima Centre in combination with the multi-center MRI database.

 

Brief summary of progress / results July 2024

This year began with the presentation of our work at the Connective Tissue Oncology Society (CTOS) conference of 2023, highlighting how apparent diffusion coefficient (ADC) maps automatically calculated by the scanner can be heterogeneous, especially when compared with ADC maps calculated using a standardized pipeline. This can create an ulterior layer of variability in multicenter study, that could be avoided by the use of a standardize pipeline for calculation of ADC maps from diffusion-weighted images.

A project towards complete 3D dataset annotation was started. For each patient, the ADC maps and at least one type of structural image (usually contrast-enhanced T1 weighted image) were manually segmented. Certified radiologist will check these images, to be used as ground truth for automatic segmentation and possibly for a more in-depth study on the shape, volume, and internal structure of the tumor.

After a period of self-study on AI, having automatic segmentation as a goal, a series of neural networks were implemented. A series of preliminary experiments on U-net and nnU-net were started in preparation for the complete and checked dataset. For both architectures, 3D and 2D options were explored, and different MRI modalities as input were compared (ADC maps vs structural). Preliminary result indicate that the use of structural images tends to give better results and a series of more structured experiments will be undertaken next year.

The postdoc started working on the harmonization task since April 2024. So far, the dataset was curated in order to minimize the bias in the observed difference between diagnostic and follow-up scans. The data in the selected subset was pre-processed by registering follow-up images to diagnostic ones as well as transforming all the masks. It was shown that at initial stage there is a significant difference between the ADC distributions of both healthy and tumor areas when comparing longitudinal scans. A baseline was obtained for harmonization using quantile normalization approach that was applied to manually annotated regions of interest that surround and include the tumor area.

 

Brief summary of progress / results July 2023

MRI scans of pediatric patients with rhabdomyosarcoma were collected from multiple centers across Europe. In the first months of this project, the images and relative metadata were organized in a database. Given our interest in harmonization of diffusion MRI data, a preliminary study was conducted trying to ascertain whether ADC maps automatically calculated from scanners from different vendors are comparable to ADC maps calculated with a standardized in-house pipeline. Values within the tumors were compared and no significant difference has been found in the subset of images analyzed. While the research to expand our current dataset using already existing scans is still ongoing, an MRI acquisition protocol for the acquisition of a multi-scanner MRI database on volunteers has been developed. Moreover, a framework for manual segmentation of tumor and main healthy structures has been implemented, to be used as ground truth in future steps of the study.