Ronald Boellaard

Amsterdam UMC, location VUMC

In short, Ronald Boellaard's research entails the following: 

'Artificial intelligence and machine learning for FDG PET/CT response prediction in diffuse large B cell lymphoma

The project aims at the application of artificial intelligence (AI) methods to improve prognosis and treatment response prediction based on FDG PET/CT studies of diffuse large B-cell lymphoma patients. This project is an extension of a successfully running international consortium project PETRA (petralymphoma.org). The Hanarth Fonds grant will be used to investigate the use of radiomics analysis in combination with machine learning as well as the use of convolution neural networks for deep radiomics analysis of FDG PET/CT to better predict response to treatment, thereby avoiding futile treatments, as compared to current FDG PET/CT reads.'

Short summary of progress / results May 2024

Last year we developed a convolutional neural network for prediction of 2 years time to progression (TTP) in diffuse large B-cell lymphoma patients. Baseline scans from 296 DLBCL patients were included in this study. The scans were converted into maximum intensity projections to generate 2D images out of the scans. This reduced the size of the data and thus the computational time. We found this CNN to be able to predict TTP with an Area under the Curve of 0.72 (compared to 0.68 achieved by the international prognostic index, currently used in the clinic). This model was externally tested on a second dataset of 340 DLBCL patients yielding a performance of 0.74. Next, we evaluated the sensitivity of the predictions of the CNN to reconstruction protocols. We found that the predictions changed depending on the reconstruction used. We investigated two methods to compensate for this effect: ComBat and image based transformation. The latter consisted in a transformation of the images by blurring them with Gaussian filters. The image transformation happened to be very effective, removing the prediction differences between the reconstructions. Most recently, we performed external testing using 6 international trial datasets and we confirmed that the deep learning model consistently outperformed the so-called IPI score.

Currently we are building a 3D CNN model which will use the whole scans (instead of the MIPs). We will compare the performance to our MIP CNN to investigate whether MIPs already contain enough information and these are an alternative to 3D data for tumor response prediction in DLBCL. Moreover, we are testing several publicly available tools for AI based PET tumor segmentation and use these segmentations as starting point in our radiomics pipeline. In the near future we intend to compare or combine deep learning based outcome prediction with this AI-segmentation based outcome prediction models and identify to most optimal approach for predicting treatment outcome.

Short summary of progress / results May 2023

Last year we developed a convolutional neural network for prediction of 2 years time to progression (TTP) in diffuse large B-cell lymphoma patients. Baseline scans from 296 DLBCL patients were included in this study. The scans were converted into maximum intensity projections to generate 2D images out of the scans. This reduced the size of the data and thus the computational time. We found this CNN to be able to predict TTP with an Area under the Curve of 0.72 (compared to 0.68 achieved by the international prognostic index, currently used in the clinic). This model was externally tested on a second dataset of 340 DLBCL patients yielding a performance of 0.74.

Next, we evaluated the sensitivity of the predictions of the CNN to reconstruction protocols. We found that the predictions changed depending on the reconstruction used. We investigated two methods to compensate for this effect: ComBat and image based transformation. The latter consisted in a transformation of the images by blurring them with Gaussian filters. The image transformation happened to be very effective, removing the prediction differences between the reconstructions.

Currently we are building a 3D CNN model which will use the whole scans (instead of the MIPs). We will compare the performance to our MIP CNN to investigate whether MIPs already contain enough information and these are an alternative to 3D data for tumor response prediction in DLBCL. In parallel, Interim PETs are being collected and preprocessed. Interim PETs are the scans obtained after 2 or 4 cycles of treatment. The aim is to generate a MIP CNN model with interim scans and see what is the added value of these to the prediction.

Short summary of progress / results May 2022

During the first 6 month of the project a data harmonisation method was succesfully implemented, applied and tested to mitigate differences in FDG PET uptake metrics due to use of different image reconstructions (as seen in multicenter studies) and tumor segmentation methods (as seen in  different softwares). We found that the method was able to harmonize data for some segmentation methods, but also that use of fixed size thresholds seem to allow extraction of features fairly independent on image reconstruction and thereby bypasses the need for retrospective data harmonization. 

Secondly, we developped a deep learning method (CNN) that used maximum intensity projection of the FDG PET/CT data and we found that the method has feasibility to predict 2 years time to progression in DLBCL patients with an externally validated ROC-AUC of 0.70. Currently, we aim to extend to model including other image derived metrics as well as clinical data to further improve performance. Moreover, explainable components will be added to the deep learning method(s) as to assist the user in assessing the plausiblity of the provided predictions.