Paul Suetens’s research while affiliated with KU Leuven and other places

What is this page?


This page lists works of an author who doesn't have a ResearchGate profile or hasn't added the works to their profile yet. It is automatically generated from public (personal) data to further our legitimate goal of comprehensive and accurate scientific recordkeeping. If you are this author and want this page removed, please let us know.

Publications (553)


Figure 2: The expected SD loss value as a function of the predicted foreground probability˜pprobability˜ probability˜p β for different true foreground probabilities p β (a higher opacity points to a higher true foreground probability p β = {0, 0.25, 0.5, 0.75, 1}). Different rows/colors represent different total volumes of the uncertain area: µ = 0.25 (blue), µ = 1 (black), µ = 4 (red). Numerical results are given for K={1, 4, 16} independent regions, respectively left to right column.
Figure 3: Error on the predicted foreground probability as a function of the true foreground probability p β after SD optimization. Different colors represent different total volumes of the uncertain area: µ = 0.25 (blue), µ = 1 (black), µ = 4 (red). Numerical results are given for K={1, 4, 16} independent regions, respectively left to right column.
Figure 4: Network architectures. LR model in dark gray rectangle. U-Net S model in light gray rectangle. For U-Net M and L models this extends up until the dashed line. Legend: f -# input features (dataset dependent)); n -# feature maps; unfilled horizontal arrow -1x1(x1) convolution with sigmoid activation; filled horizontal arrow -3x3(x3) convolution with leaky-ReLU activation; down arrow -average/max-pooling; up arrow -bi/tri-linear upsampling; dashed horizontal arrow: alignment cropping and concatenation.
Theoretical analysis and experimental validation of volume bias of soft Dice optimized segmentation maps in the context of inherent uncertainty
  • Preprint
  • File available

November 2022

·

32 Reads

·

·

·

Paul Suetens

The clinical interest is often to measure the volume of a structure, which is typically derived from a segmentation. In order to evaluate and compare segmentation methods, the similarity between a segmentation and a predefined ground truth is measured using popular discrete metrics, such as the Dice score. Recent segmentation methods use a differentiable surrogate metric, such as soft Dice, as part of the loss function during the learning phase. In this work, we first briefly describe how to derive volume estimates from a segmentation that is, potentially, inherently uncertain or ambiguous. This is followed by a theoretical analysis and an experimental validation linking the inherent uncertainty to common loss functions for training CNNs, namely cross-entropy and soft Dice. We find that, even though soft Dice optimization leads to an improved performance with respect to the Dice score and other measures, it may introduce a volume bias for tasks with high inherent uncertainty. These findings indicate some of the method's clinical limitations and suggest doing a closer ad-hoc volume analysis with an optional re-calibration step.

Download

Figure 1 Schematic of workflow. (A) Global-to-local segmentation with visualization of the 1st to 6th level of the hierarchical segmentation
Figure 2 Asymptomatic carriers C9orf72 versus non-carriers. Global-to-local segment results for A size and B shape and their respective dendrograms. Asterisk indicates FDR-adjusted significance (dep) P = 0.0007, −log P = 3.17, results below the FDR-adjusted significance threshold are not illustrated. Nodes can be linked to their spatial coverage via Fig. 1A. Results from other levels are shown in Supplementary Figure 3. (C) VBM and TBM analysis: asterisk indicates FDR-adjusted significance (dep) P = 0.0003, t = 3.46. Note that the by convention, only atrophy is visualized so widening of the ventricles cannot be assessed from the univariate results.
Figure 3 Symptomatic carriers C9orf72 versus non-carriers. Global-to-local segment results for A size and B shape and their respective dendrograms. Asterisk indicates FDR-adjusted significance (dep) P = 0.0007, −log P = 3.17, results below the FDR-adjusted significance threshold are not illustrated. Nodes can be linked to their spatial coverage via Fig. 1A. Results from other levels are shown in Supplementary Figure 4. (C) VBM and TBM analysis: asterisk indicates FDR-adjusted significance (dep) P = 0.0003, t = 3.46.
Clinical information of the patients of the GENFI cohort in this study
Hierarchical spectral clustering reveals brain size and shape changes in asymptomatic carriers of C9orf72

July 2022

·

164 Reads

·

2 Citations

Brain Communications

·

Dorothy Gors

·

Alicia Bárcenas Gallardo

·

[...]

·

Traditional methods for detecting asymptomatic brain changes in neurodegenerative diseases such as Alzheimer’s disease or frontotemporal degeneration (FTD) typically evaluate changes in volume at a predefined level of granularity, e.g. voxel-wise or in a priori defined cortical volumes of interest. Here we apply a method based on hierarchical spectral clustering, a graph-based partitioning technique. Our method uses multiple levels of segmentation for detecting changes in a data-driven, unbiased, comprehensive manner within a standard statistical framework. Furthermore, spectral clustering allows for detection of changes in shape along with changes in size. We performed tensor-based morphometry to detect changes in the Genetic Frontotemporal dementia Initiative (GENFI) asymptomatic and symptomatic FTD mutation carriers using hierarchical spectral clustering and compared the outcome to that obtained with a more conventional voxel-wise tensor- and voxel-based morphometric analysis. In the symptomatic groups, the hierarchical spectral clustering-based method yielded results that were largely in line with those obtained with the voxel-wise approach. In asymptomatic C9orf72 expansion carriers, spectral clustering detected changes in size in medial temporal cortex that voxel-wise methods could only detect in the symptomatic phase. Furthermore, in the asymptomatic and the symptomatic phases, the spectral clustering approach detected changes in shape in the premotor cortex in C9orf72. In summary, the current study shows the merit of hierarchical spectral clustering for data-driven segmentation and detection of structural changes in the symptomatic and asymptomatic stages of monogenic FTD.


MeVisLab-OpenVR prototyping platform for virtual reality medical applications

June 2022

·

40 Reads

·

5 Citations

International Journal of Computer Assisted Radiology and Surgery

Purpose: Virtual reality (VR) can provide an added value for diagnosis and/or intervention planning. Several VR software implementations have been proposed but they are often application dependent. Previous attempts for a more generic solution incorporating VR in medical prototyping software (MeVisLab) were still lacking functionality precluding easy and flexible development. Methods: We propose an alternative solution that uses rendering to a graphical processing unit (GPU) texture to enable rendering arbitrary Open Inventor scenes in a VR context. It facilitates flexible development of user interaction and rendering of more complex scenes involving multiple objects. We tested the platform in planning a transcatheter cardiac stent placement procedure. Results: This approach proved to enable development of a particular implementation that facilitates planning of percutaneous treatment of a sinus venosus atrial septal defect. The implementation showed it is intuitive to plan and verify the procedure using VR. Conclusion: An alternative implementation for linking OpenVR with MeVisLab is provided that offers more flexible development of VR prototypes which can facilitate further clinical validation of this technology in various medical disciplines.


Prediction of Stroke Infarct Growth Rates by Baseline Perfusion Imaging

September 2021

·

49 Reads

·

26 Citations

Stroke

Background and Purpose Computed tomography perfusion imaging allows estimation of tissue status in patients with acute ischemic stroke. We aimed to improve prediction of the final infarct and individual infarct growth rates using a deep learning approach. Methods We trained a deep neural network to predict the final infarct volume in patients with acute stroke presenting with large vessel occlusions based on the native computed tomography perfusion images, time to reperfusion and reperfusion status in a derivation cohort (MR CLEAN trial [Multicenter Randomized Clinical Trial of Endovascular Treatment for Acute Ischemic Stroke in the Netherlands]). The model was internally validated in a 5-fold cross-validation and externally in an independent dataset (CRISP study [CT Perfusion to Predict Response to Recanalization in Ischemic Stroke Project]). We calculated the mean absolute difference between the predictions of the deep learning model and the final infarct volume versus the mean absolute difference between computed tomography perfusion imaging processing by RAPID software (iSchemaView, Menlo Park, CA) and the final infarct volume. Next, we determined infarct growth rates for every patient. Results We included 127 patients from the MR CLEAN (derivation) and 101 patients of the CRISP study (validation). The deep learning model improved final infarct volume prediction compared with the RAPID software in both the derivation, mean absolute difference 34.5 versus 52.4 mL, and validation cohort, 41.2 versus 52.4 mL ( P <0.01). We obtained individual infarct growth rates enabling the estimation of final infarct volume based on time and grade of reperfusion. Conclusions We validated a deep learning-based method which improved final infarct volume estimations compared with classic computed tomography perfusion imaging processing. In addition, the deep learning model predicted individual infarct growth rates which could enable the introduction of tissue clocks during the management of acute stroke.


Abstract P333: Prediction of Stroke Lesion Growth Rates by Baseline Perfusion Imaging

March 2021

·

18 Reads

·

4 Citations

Stroke

Objective: Computed Tomography Perfusion imaging (CTP) allows estimation of tissue status in patients with acute ischemic stroke. We aimed to improve prediction of the final infarct and individual infarct growth rates based on a deep learning approach. Methods: We trained a deep neural network to predict the final infarct volume in patients presenting with large vessel occlusions based on the native CTP images, time to reperfusion and reperfusion status in a derivation cohort (MR CLEAN study). The model was internally validated in a five-fold cross-validation and externally in an independent dataset (CRISP study). We calculated the mean absolute difference (MAD) between the predictions of the deep learning model and the final infarct volume versus the MAD between CTP processing by RAPID software and the final infarct volume. Next, we determined infarct growth rates for every patient. Results: We included 127 patients from the MR CLEAN (derivation) and 101 patients of the CRISP study (validation). The deep learning model improved final infarct lesion prediction compared to the RAPID software in both the derivation, MAD 34.5 vs 52.4ml, and validation cohort, 41.2 vs 52.4 ml, (p < 0.01). We obtained individual infarct growth rates enabling the estimation of final infarct volume based on time and grade of reperfusion. Interpretation: We validated a deep learning-based method which improved final infarct volume estimations compared to classic CTP processing. In addition, the deep learning model predicted individual infarct growth rates which could potentially enable the introduction of tissue clocks during the management of acute stroke. Figure A. Patient with a mean infarct growth of 18.3 ml/h. The final infarct volume was 104 ml. Recanalization was performed 131 min after CT perfusion with a mTICI = 2b. B . Patient with a mean infarct growth of 2.3 ml/h. The final infarct volume was 10.8 ml. Recanalization was performed 101 min after CT perfusion with a mTICI = 3.


Explainable-by-design Semi-Supervised Representation Learning for COVID-19 Diagnosis from CT Imaging

November 2020

·

529 Reads

Our motivating application is a real-world problem: COVID-19 classification from CT imaging, for which we present an explainable Deep Learning approach based on a semi-supervised classification pipeline that employs variational autoencoders to extract efficient feature embedding. We have optimized the architecture of two different networks for CT images: (i) a novel conditional variational autoencoder (CVAE) with a specific architecture that integrates the class labels inside the encoder layers and uses side information with shared attention layers for the encoder, which make the most of the contextual clues for representation learning, and (ii) a downstream convolutional neural network for supervised classification using the encoder structure of the CVAE. With the explainable classification results, the proposed diagnosis system is very effective for COVID-19 classification. Based on the promising results obtained qualitatively and quantitatively, we envisage a wide deployment of our developed technique in large-scale clinical studies.Code is available at https://git.etrovub.be/AVSP/ct-based-covid-19-diagnostic-tool.git.


Theoretical analysis and experimental validation of volume bias of soft Dice optimized segmentation maps in the context of inherent uncertainty

September 2020

·

55 Reads

·

24 Citations

Medical Image Analysis

The clinical interest is often to measure the volume of a structure, which is typically derived from a segmentation. In order to evaluate and compare segmentation methods, the similarity between a seg-mentation and a predefined ground truth is measured using popular discrete metrics, such as the Dice score. Recent segmentation methods use a differentiable surrogate metric, such as soft Dice, as part of the loss function during the learning phase. In this work, we first briefly describe how to derive volume estimates from a segmentation that is, potentially, inherently uncertain or ambiguous. This is followed by a theoretical analysis and an experimental validation linking the inherent uncertainty to common loss functions for training CNNs, namely cross-entropy and soft Dice. We find that, even though soft Dice optimization leads to an improved performance with respect to the Dice score and other measures, it may introduce a volume bias for tasks with high inherent uncertainty. These findings indicate some of the method's clinical limitations and suggest doing a closer ad-hoc volume analysis with an optional re-calibration step.


Comparative study of deep learning methods for the automatic segmentation of lung, lesion and lesion type in CT scans of COVID-19 patients

July 2020

·

346 Reads

Recent research on COVID-19 suggests that CT imaging provides useful information to assess disease progression and assist diagnosis, in addition to help understanding the disease. There is an increasing number of studies that propose to use deep learning to provide fast and accurate quantification of COVID-19 using chest CT scans. The main tasks of interest are the automatic segmentation of lung and lung lesions in chest CT scans of confirmed or suspected COVID-19 patients. In this study, we compare twelve deep learning algorithms using a multi-center dataset, including both open-source and in-house developed algorithms. Results show that ensembling different methods can boost the overall test set performance for lung segmentation, binary lesion segmentation and multiclass lesion segmentation, resulting in mean Dice scores of 0.982, 0.724 and 0.469, respectively. The resulting binary lesions were segmented with a mean absolute volume error of 91.3 ml. In general, the task of distinguishing different lesion types was more difficult, with a mean absolute volume difference of 152 ml and mean Dice scores of 0.369 and 0.523 for consolidation and ground glass opacity, respectively. All methods perform binary lesion segmentation with an average volume error that is better than visual assessment by human raters, suggesting these methods are mature enough for a large-scale evaluation for use in clinical practice.



Optimization with Soft Dice Can Lead to a Volumetric Bias

May 2020

·

55 Reads

·

27 Citations

Lecture Notes in Computer Science

Segmentation is a fundamental task in medical image analysis. The clinical interest is often to measure the volume of a structure. To evaluate and compare segmentation methods, the similarity between a segmentation and a predefined ground truth is measured using metrics such as the Dice score. Recent segmentation methods based on convolutional neural networks use a differentiable surrogate of the Dice score, such as soft Dice, explicitly as the loss function during the learning phase. Even though this approach leads to improved Dice scores, we find that, both theoretically and empirically on four medical tasks, it can introduce a volumetric bias for tasks with high inherent uncertainty. As such, this may limit the method’s clinical applicability.


Citations (70)


... Hierarchical Spectral Clustering, a graph-based partitioning technique, has been showing the merit of datadriven segmentation (Brandi and Di Matteo 2021, Wang et al. 2022, Yang et al. 2020. It is characterized by strong migration of nodes and refers to spectral clustering at a continuous level to segment the data into increasingly finer levels of detail, starting from a broad, global segmentation and moving towards more specific, local segmentations (Bruffaerts et al. 2022). In addition, it is coupled with a comprehensive statistical approach that could identify clusters with different distributions of the samples. ...

Reference:

Integrating physics-based fragility for hierarchical spectral clustering for resilience assessment of power distribution systems under extreme winds
Hierarchical spectral clustering reveals brain size and shape changes in asymptomatic carriers of C9orf72

Brain Communications

... As industrial and academic research groups regularly embark on ambitious biomedical R &D projects, numerous software has been developed to support their efforts. Fortunately, some of them are freely distributed to the research community [1][2][3][4][5][6][7][8][9][10][11][12][13][14][15][16][17][18][19]. Due to the lack of standardization and comparability among these software solutions, researchers often struggle to identify the most suitable software for their needs. ...

MeVisLab-OpenVR prototyping platform for virtual reality medical applications
  • Citing Article
  • June 2022

International Journal of Computer Assisted Radiology and Surgery

... It also leads to limited or no means of external validation which creates a significant obstacle to clinical translation [141]. Such lack of external validation, which has been explored by only a few [33,85,101,124], is cause for concern on the reliability of reported results. ...

Prediction of Stroke Infarct Growth Rates by Baseline Perfusion Imaging
  • Citing Article
  • September 2021

Stroke

... With the input of a particular reperfusion status at a particular timepoint after imaging, the model was able to estimate the corresponding final infarct volume. Besides, the evolution of the infarct would be presented in a movie and the mean lesion growth rate per hour could be calculated [61]. It provided a new way to demonstrate the change of infarct core individually with time flying. ...

Abstract P333: Prediction of Stroke Lesion Growth Rates by Baseline Perfusion Imaging
  • Citing Article
  • March 2021

Stroke

... Especially, the PhiSeg, a hierarchical probabilistic UNet, achieved outstanding calibration scores and uncertainty-error overlap, although it demonstrated lower segmentation performance (DSC: 0.73, p < 0.05) for GTV-T compared to the Baseline (DSC: 0.75). Except for PhiSeg, all methods utilised a hybrid loss function that combined cross-entropy and Dice loss, with the latter aiding small target segmentation but potentially resulting in overconfident segmentations (Bertels et al 2021). For example, figure 4 displayed a more blurred edge on the probability map and a thicker uncertainty band on the GTV-T for PhiSeg, with a smoother gradient transitioning from highly uncertain to low uncertain regions. ...

Theoretical analysis and experimental validation of volume bias of soft Dice optimized segmentation maps in the context of inherent uncertainty
  • Citing Article
  • September 2020

Medical Image Analysis

... Since polyp size is very often overestimated, Suykens et al. reported an AI system that can objectively infer polyp size based on a reference tool (i.e. biopsy forceps) in the endoscopic image [94]. They used two separate deep learning algorithms: (1) delineation of the polyp and (2) detection of two landmarks on the forceps, resulting in a size estimation that can detect the polyp and forceps in 71% of the 35 test images and a decrease in overestimation bias by 63% (p-value < 0.1). ...

Sa2012 AUTOMATED POLYP SIZE ESTIMATION WITH DEEP LEARNING REDUCES INTEROBSERVER VARIABILITY

Gastrointestinal Endoscopy

... In the last decade, many works proposed computationally efficient implementations based on the primal-dual optimization algorithm of Chambolle and Pock [22] for dynamic reconstructions [27]- [29]. Recently we have proposed a dynamical iodine reconstruction based on a 2step method with data acquired with dual-energy (DE) CBCT devices [30] with a motion-correction extension [31]. ...

4D CBCT reconstruction with TV regularization on a dynamic software phantom
  • Citing Conference Paper
  • October 2019

... The suggested network model yielded impressive results and has significant value in clinical breast-imaging applications. De Buck et al. [189] proposed a novel approach that integrates artificial deep learning-based breast segmentation from CT thorax exams with radiodensity and volumetric breast densitybased breast glandular computation. A cutting-edge CNN was trained to segment the breast area, allowing them to compute BrC risk scores on the segmented CT volume. ...

Automated breast cancer risk estimation on routine CT thorax scans by deep learning segmentation
  • Citing Conference Paper
  • March 2020

... Robben et al. [78] developed a CNN model, based on DeepMedic [96], using spatiotemporal CTP data and clinical information to predict follow-up infarct lesions, as shown in Fig. 2. The model architecture comprises four branches that extract the features of CTP images and their downsampled versions, arterial input function, and metadata. These features are concatenated and fed through three convolutional layers to generate the final infarct segmentation map. ...

Prediction of final infarct volume from native CT perfusion and treatment parameters using deep learning
  • Citing Article
  • October 2019

Medical Image Analysis