Joren Brunekreef
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Joren Brunekreef has a total of 48 co-authors. Here are the top co-authors by number of shared publications:
- Jonas Teuwen
Department of radiation oncology, Netherlands Cancer Institute, Amsterdam, The Netherlands
5 shared publications - Renate Loll
Radboud University
4 shared publications - Eric Marcus
AI for Oncology Netherlands Cancer Institute Amsterdam the Netherlands
3 shared publications - Hendrik Nunner
Institute of Human-Centered Interactive Systems (IMIS), University of Lübeck, Ratzeburger Allee 160, Lübeck, 23562, Germany, 49 451-3101 5117
2 shared publications - Lishan Cai
Department of Radiology, The Netherlands Cancer Institute, Amsterdam, The Netherlands
2 shared publications - Mirjam Kretzschmar
Julius Center for Health Sciences and Primary care, University Medical Center Utrecht, Utrecht University, Utrecht, the Netherlands
2 shared publications
Latest Publications
- Open Access
Predicting Rectal Cancer Patient Survival with Dutch Radiology Reports using Natural Language Processing (NLP): The Role of Pretrained Language Models
Summary: The use of Electronic Health Records (EHRs) has increased significantly in recent years. However, a substantial portion of the clinical data remains in unstructured text formats, especially in the context of radiology. This limits the application of EHRs for automated analysis in oncology research. Pretrained language models have been utilized to extract feature embeddings from these reports for downstream clinical applications, such as treatment response and survival prediction. However, a thorough investigation into which pretrained models produce the most effective features for rectal cancer survival prediction has not yet been done. This study explores the performance of five Dutch pretrained language models, including two publicly available models (RobBERT and MedRoBERTa.nl) and three developed in-house for the purpose of this study (RecRoBERT, BRecRoBERT, and BRec2RoBERT) with training on distinct Dutch-only corpora, in predicting overall survival and disease-free survival outcomes in rectal cancer patients. Our results showed that our in-house developed BRecRoBERT, a RoBERTa-based language model trained from scratch on a combination of Dutch breast and rectal cancer corpora, delivered the best predictive performance for both survival tasks, achieving a C-index of 0.65 (0.57, 0.73) for overall survival and 0.71 (0.64, 0.78) for disease-free survival. It outperformed models trained on general Dutch corpora (RobBERT) or Dutch hospital clinical notes (MedRoBERTa.nl). BRecRoBERT demonstrated the potential capability to predict survival in rectal cancer patients using Dutch radiology reports at diagnosis. This study highlights the value of pretrained language models that incorporate domain-specific knowledge for downstream clinical applications. Furthermore, it proves that utilizing data from related domains can improve the quality of feature embeddings for certain clinical tasks, particularly in situations where domain-specific data is scarce.
Open AccessPredicting Rectal Cancer Patient Survival with Dutch Radiology Reports using Natural Language Processing (NLP): The Role of Pretrained Language Models
Summary: The use of Electronic Health Records (EHRs) has increased significantly in recent years. However, a substantial portion of the clinical data remains in unstructured text formats, especially in the context of radiology. This limits the application of EHRs for automated analysis in oncology research. Pretrained language models have been utilized to extract feature embeddings from these reports for downstream clinical applications, such as treatment response and survival prediction. However, a thorough investigation into which pretrained models produce the most effective features for rectal cancer survival prediction has not yet been done. This study explores the performance of five Dutch pretrained language models, including two publicly available models (RobBERT and MedRoBERTa.nl) and three developed in-house for the purpose of this study (RecRoBERT, BRecRoBERT, and BRec2RoBERT) with training on distinct Dutch-only corpora, in predicting overall survival and disease-free survival outcomes in rectal cancer patients. Our results showed that our in-house developed BRecRoBERT, a RoBERTa-based language model trained from scratch on a combination of Dutch breast and rectal cancer corpora, delivered the best predictive performance for both survival tasks, achieving a C-index of 0.65 (0.57, 0.73) for overall survival and 0.71 (0.64, 0.78) for disease-free survival. It outperformed models trained on general Dutch corpora (RobBERT) or Dutch hospital clinical notes (MedRoBERTa.nl). BRecRoBERT demonstrated the potential capability to predict survival in rectal cancer patients using Dutch radiology reports at diagnosis. This study highlights the value of pretrained language models that incorporate domain-specific knowledge for downstream clinical applications. Furthermore, it proves that utilizing data from related domains can improve the quality of feature embeddings for certain clinical tasks, particularly in situations where domain-specific data is scarce.
- Open Access
- 19 December 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (12) , e0338818
Impact of adherence and stringency on the effectiveness of lockdown measures: A modelling study
During the COVID-19 pandemic, lockdowns were a widely used strategy to reduce disease transmission. However, there was much debate about the optimal level of strictness and duration of lockdowns. This study considers how lockdowns impact public health opinions, which in turn influence adherence to and effectiveness of these measures. We developed an agent-based simulation model to theoretically explore the impact of health-related opinions on the effectiveness of lockdowns in controlling disease spread. We simulated these dynamics within a hypothetical population connected via a simplified contact network (Watts-Strogatz), incorporating feedback loops between disease prevalence, opinions, and behaviour, including ‘lockdown fatigue’. We explored different scenarios of lockdown implementation in our hypothetical population network by varying a threshold value of prevalence when a lockdown is initiated and the stringency of the lockdown. Our qualitative findings imply that quickly imposing a lockdown with high stringency is the most effective at reducing infection spread, provided that there is a certain degree of adherence to the lockdown among the population. Furthermore, stricter lockdowns minimize fatigue with respect to the imposed measures, since the duration of a lockdown is shorter on average in this scenario. These theoretical results imply that such lockdown policies might therefore be a beneficial, high-impact tool in containing epidemic spread, especially when supplemented by information interventions maintaining the adherence to lockdown measures.
- 19 December 2025
- Published by Public Library of Science (PLoS) in PLOS ONE
- Vol. 20 (12) , e0338818
Open AccessImpact of adherence and stringency on the effectiveness of lockdown measures: A modelling study
During the COVID-19 pandemic, lockdowns were a widely used strategy to reduce disease transmission. However, there was much debate about the optimal level of strictness and duration of lockdowns. This study considers how lockdowns impact public health opinions, which in turn influence adherence to and effectiveness of these measures. We developed an agent-based simulation model to theoretically explore the impact of health-related opinions on the effectiveness of lockdowns in controlling disease spread. We simulated these dynamics within a hypothetical population connected via a simplified contact network (Watts-Strogatz), incorporating feedback loops between disease prevalence, opinions, and behaviour, including ‘lockdown fatigue’. We explored different scenarios of lockdown implementation in our hypothetical population network by varying a threshold value of prevalence when a lockdown is initiated and the stringency of the lockdown. Our qualitative findings imply that quickly imposing a lockdown with high stringency is the most effective at reducing infection spread, provided that there is a certain degree of adherence to the lockdown among the population. Furthermore, stricter lockdowns minimize fatigue with respect to the imposed measures, since the duration of a lockdown is shorter on average in this scenario. These theoretical results imply that such lockdown policies might therefore be a beneficial, high-impact tool in containing epidemic spread, especially when supplemented by information interventions maintaining the adherence to lockdown measures.
- Open Access
Integrating computational protein structure predictions and genetic dependencies yields an atlas of human multi-protein complexes (AHMPC)
Knowledge of which proteins interact to form functional complexes in cells is essential for understanding molecular mechanisms in biology. Structure prediction methods recently allowed to compute the Human Interactome of likely binary protein interactions. We combine computational predictions with orthogonal functional data from the Dependency Map that estimate the correlation between genetic dependencies and vulnerabilities of the corresponding gene pairs. This revealed groups of proteins that likely form larger complexes. Clustering analysis followed by AlphaFold3 multi-protein complex predictions and AlphaBridge analysis provided the basis to construct and atlas of human multi-protein complexes (AHMPC), currently encompassing 354 high-confidence predicted multi-protein complexes. These include well-known assemblies and new ones - such as a complex involving SYS1, JTB, and ARFRP1 that we validate experimentally, suggesting an unexpected role of JTB in Golgi traficking. To enable the research community to explore the AHMPC and enable further discovery, we cluster all complexes using functional and disease-related embeddings, demonstrate how structured prompts allow validation by large language models (LLMs), and make all structures and analysis available online as an open resource at https://ahmpc.eu/ .
Open AccessIntegrating computational protein structure predictions and genetic dependencies yields an atlas of human multi-protein complexes (AHMPC)
Knowledge of which proteins interact to form functional complexes in cells is essential for understanding molecular mechanisms in biology. Structure prediction methods recently allowed to compute the Human Interactome of likely binary protein interactions. We combine computational predictions with orthogonal functional data from the Dependency Map that estimate the correlation between genetic dependencies and vulnerabilities of the corresponding gene pairs. This revealed groups of proteins that likely form larger complexes. Clustering analysis followed by AlphaFold3 multi-protein complex predictions and AlphaBridge analysis provided the basis to construct and atlas of human multi-protein complexes (AHMPC), currently encompassing 354 high-confidence predicted multi-protein complexes. These include well-known assemblies and new ones - such as a complex involving SYS1, JTB, and ARFRP1 that we validate experimentally, suggesting an unexpected role of JTB in Golgi traficking. To enable the research community to explore the AHMPC and enable further discovery, we cluster all complexes using functional and disease-related embeddings, demonstrate how structured prompts allow validation by large language models (LLMs), and make all structures and analysis available online as an open resource at https://ahmpc.eu/ .
- Open Access
A quantitative comparison between human experts and AI at estimating tumor-stroma ratio
The tumor–stroma ratio (TSR) is an established prognostic biomarker across several cancer types, yet its manual assessment remains labour-intensive and subject to inter-observer variability. An artificial intelligence (AI)-based estimate could offer an efficient, consistent alternative for this task. In this study, quantitative comparisons were made between expert humans and an AI model for TSR estimation. Using two independent, multi-institutional histopathology datasets, an Attention U-Net was benchmarked against experienced pathologists. In a subset of the TCGA-BRCA dataset, the AI model demonstrated comparable trends to human consensus for TSR quantification, achieving an intraclass correlation coefficient (ICC) of 0.69. However, the AI model’s TSR scores are on average 5 percentage points higher compared to human scores on this dataset. The AI model was found to be more consistent at estimating TSR than either of the human counterparts, with a discrepancy ratio (DR) of 0.86. Results on an external dataset obtained from the Netherlands Cancer Institute consisting of cases (n=357) from 35 different Dutch hospitals showed that the AI model’s TSR scores are 7 percentage points lower on average compared to the human rater with an ICC of 0.59. To account for the model’s imperfect segmentation performance, we derived an estimate of the ambiguity in AI-based TSR predictions. The results indicate that, despite this ambiguity, the AI not only follows similar trends but also delivers greater overall scoring consistency than manual TSR assessment by humans.
Open AccessA quantitative comparison between human experts and AI at estimating tumor-stroma ratio
The tumor–stroma ratio (TSR) is an established prognostic biomarker across several cancer types, yet its manual assessment remains labour-intensive and subject to inter-observer variability. An artificial intelligence (AI)-based estimate could offer an efficient, consistent alternative for this task. In this study, quantitative comparisons were made between expert humans and an AI model for TSR estimation. Using two independent, multi-institutional histopathology datasets, an Attention U-Net was benchmarked against experienced pathologists. In a subset of the TCGA-BRCA dataset, the AI model demonstrated comparable trends to human consensus for TSR quantification, achieving an intraclass correlation coefficient (ICC) of 0.69. However, the AI model’s TSR scores are on average 5 percentage points higher compared to human scores on this dataset. The AI model was found to be more consistent at estimating TSR than either of the human counterparts, with a discrepancy ratio (DR) of 0.86. Results on an external dataset obtained from the Netherlands Cancer Institute consisting of cases (n=357) from 35 different Dutch hospitals showed that the AI model’s TSR scores are 7 percentage points lower on average compared to the human rater with an ICC of 0.59. To account for the model’s imperfect segmentation performance, we derived an estimate of the ambiguity in AI-based TSR predictions. The results indicate that, despite this ambiguity, the AI not only follows similar trends but also delivers greater overall scoring consistency than manual TSR assessment by humans.
- Open Access
Improving Rectal Tumor Segmentation with Anomaly Fusion Derived from Anatomical Inpainting: A Multicenter Study
Accurate rectal tumor segmentation using magnetic resonance imaging (MRI) is paramount for effective treatment planning. It allows for volumetric and other quantitative tumor assessments, potentially aiding in prognostication and treatment response evaluation. Manual delineation of rectal tumors and surrounding structures is time-consuming and typically. Over the past few years, deep learning has shown strong results in automated tumor segmentation in MRI. Current studies on automated rectal tumor segmentation, however, focus solely on tumoral regions without considering the rectal anatomical entities and often lack a solid multicenter external validation. In this study, we improved rectal tumor segmentation by incorporating anomaly maps derived from anatomical inpainting. This inpainting was implemented using a U-Net-based model trained to reconstruct a healthy rectum and mesorectum from prostate T2-weighted images (T2WI). The rectal anomaly maps were generated from the difference between the original rectal and reconstructed pseudo-healthy slices during inference. The derived anomaly maps were used in the downstream tumor segmentation tasks by fusing them as an additional input channel (AAnnUNet). Alternative methods for integrating rectal anatomical knowledge were evaluated as baselines, including Multi-Target nnUNet (MTnnUNet), which added rectum and mesorectum segmentation as auxiliary tasks, and Multi-Channel nnUNet (MCnnUNet), which utilized rectum and mesorectum masks as an additional input channel. As part of this study, we benchmarked nine models for rectal tumor segmentation on a large multicenter dataset of preoperative T2WI as the baseline and nnUNet outperformed the other eight models on the external dataset. The MTnnUNet demonstrated improvements in both supervised and semi-supervised settings (AI-generated rectum and mesoretum were used) compared to nnUNet, while the MCnnUNet showed benefits only in the semi-supervised setting. Importantly, anomaly maps were strongly associated with tumoral regions, and their integration within AAnnUNet led to the best tumor segmentation results across both settings. The effectiveness of AAnnUNet demonstrated the value of the anomaly maps, indicating a promising direction for improving rectal tumor segmentation and model robustness for multicenter data.
Open AccessImproving Rectal Tumor Segmentation with Anomaly Fusion Derived from Anatomical Inpainting: A Multicenter Study
Accurate rectal tumor segmentation using magnetic resonance imaging (MRI) is paramount for effective treatment planning. It allows for volumetric and other quantitative tumor assessments, potentially aiding in prognostication and treatment response evaluation. Manual delineation of rectal tumors and surrounding structures is time-consuming and typically. Over the past few years, deep learning has shown strong results in automated tumor segmentation in MRI. Current studies on automated rectal tumor segmentation, however, focus solely on tumoral regions without considering the rectal anatomical entities and often lack a solid multicenter external validation. In this study, we improved rectal tumor segmentation by incorporating anomaly maps derived from anatomical inpainting. This inpainting was implemented using a U-Net-based model trained to reconstruct a healthy rectum and mesorectum from prostate T2-weighted images (T2WI). The rectal anomaly maps were generated from the difference between the original rectal and reconstructed pseudo-healthy slices during inference. The derived anomaly maps were used in the downstream tumor segmentation tasks by fusing them as an additional input channel (AAnnUNet). Alternative methods for integrating rectal anatomical knowledge were evaluated as baselines, including Multi-Target nnUNet (MTnnUNet), which added rectum and mesorectum segmentation as auxiliary tasks, and Multi-Channel nnUNet (MCnnUNet), which utilized rectum and mesorectum masks as an additional input channel. As part of this study, we benchmarked nine models for rectal tumor segmentation on a large multicenter dataset of preoperative T2WI as the baseline and nnUNet outperformed the other eight models on the external dataset. The MTnnUNet demonstrated improvements in both supervised and semi-supervised settings (AI-generated rectum and mesoretum were used) compared to nnUNet, while the MCnnUNet showed benefits only in the semi-supervised setting. Importantly, anomaly maps were strongly associated with tumoral regions, and their integration within AAnnUNet led to the best tumor segmentation results across both settings. The effectiveness of AAnnUNet demonstrated the value of the anomaly maps, indicating a promising direction for improving rectal tumor segmentation and model robustness for multicenter data.
- Open Access
Impact of adherence and stringency on the effectiveness of lockdown measures: a modelling study
During the COVID-19 pandemic, lockdowns were a widely used strategy to reduce disease transmission. However, there was much debate about the optimal level of strictness and duration of lockdowns. This study considers how lockdowns impact public health opinions, which in turn influence adherence to and effectiveness of these measures. We developed an agent-based simulation model to study the impact of health-related opinions on the effectiveness of lockdowns in controlling disease spread. The model simulates infection spread and health opinion dynamics among individuals connected through a network of contacts and incorporates feedback loops between epidemic variables and opinion dynamics. We simulated different scenarios of lockdown implementation by varying a threshold value of prevalence when a lockdown is initiated and the stringency of the lockdown. We found that quickly imposing a lockdown with high stringency is the most effective at reducing infection spread, provided that there is a certain degree of adherence to the lockdown among the population. Furthermore, stricter lockdowns minimize fatigue with respect to the imposed measures, since the duration of a lockdown is shorter on average in this scenario. Such lockdown policies can therefore be a beneficial, high-impact tool in containing epidemic spread, especially when supplemented by information interventions maintaining the adherence to lockdown measures. Author summary Lockdowns were a widely used strategy to curb disease transmission during the COVID-19 pandemic. However, the effectiveness of such measures depends on the population’s adherence to the regulations. Long and strict lockdowns may lower this adherence as the population experiences ‘fatigue’ with the regulations. In our study, we model the interplay between lockdowns, disease transmission, and health-related opinions in a population network. We tested various lockdown scenarios by altering strictness of measures on the one hand, and lockdown initiation times based on disease prevalence rates on the other hand. Our findings show that swift, stringent lockdowns are the most effective in reducing infections, particularly with strong public compliance. Additionally, stricter lockdowns tend to reduce fatigue as they are generally shorter. These results highlight that prompt, rigorous lockdown policies, supported by strategies to maintain public adherence, are most efficient in controlling epidemics. Furthermore, effective communication to ensure community cooperation may enhance the success of these measures.
Open AccessImpact of adherence and stringency on the effectiveness of lockdown measures: a modelling study
During the COVID-19 pandemic, lockdowns were a widely used strategy to reduce disease transmission. However, there was much debate about the optimal level of strictness and duration of lockdowns. This study considers how lockdowns impact public health opinions, which in turn influence adherence to and effectiveness of these measures. We developed an agent-based simulation model to study the impact of health-related opinions on the effectiveness of lockdowns in controlling disease spread. The model simulates infection spread and health opinion dynamics among individuals connected through a network of contacts and incorporates feedback loops between epidemic variables and opinion dynamics. We simulated different scenarios of lockdown implementation by varying a threshold value of prevalence when a lockdown is initiated and the stringency of the lockdown. We found that quickly imposing a lockdown with high stringency is the most effective at reducing infection spread, provided that there is a certain degree of adherence to the lockdown among the population. Furthermore, stricter lockdowns minimize fatigue with respect to the imposed measures, since the duration of a lockdown is shorter on average in this scenario. Such lockdown policies can therefore be a beneficial, high-impact tool in containing epidemic spread, especially when supplemented by information interventions maintaining the adherence to lockdown measures. Author summary Lockdowns were a widely used strategy to curb disease transmission during the COVID-19 pandemic. However, the effectiveness of such measures depends on the population’s adherence to the regulations. Long and strict lockdowns may lower this adherence as the population experiences ‘fatigue’ with the regulations. In our study, we model the interplay between lockdowns, disease transmission, and health-related opinions in a population network. We tested various lockdown scenarios by altering strictness of measures on the one hand, and lockdown initiation times based on disease prevalence rates on the other hand. Our findings show that swift, stringent lockdowns are the most effective in reducing infections, particularly with strong public compliance. Additionally, stricter lockdowns tend to reduce fatigue as they are generally shorter. These results highlight that prompt, rigorous lockdown policies, supported by strategies to maintain public adherence, are most efficient in controlling epidemics. Furthermore, effective communication to ensure community cooperation may enhance the success of these measures.
- 30 April 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 37 (5) , 2706-2708
Letter to the Editor Regarding Article “Prior to Initiation of Chemotherapy, Can We Predict Breast Tumor Response? Deep Learning Convolutional Neural Networks Approach Using a Breast MRI Tumor Dataset”
The cited article reports on a convolutional neural network trained to predict response to neoadjuvant chemotherapy from pre-treatment breast MRI scans. The proposed algorithm attains impressive performance on the test dataset with a mean Area Under the Receiver-Operating Characteristic curve of 0.98 and a mean accuracy of 88%. In this letter, I raise concerns that the reported results can be explained by inadvertent data leakage between training and test datasets. More precisely, I conjecture that the random split of the full dataset in training and test sets did not occur on a patient level, but rather on the level of 2D MRI slices. This allows the neural network to “memorize” a patient’s anatomy and their treatment outcome, as opposed to discovering useful features for treatment response prediction. To provide evidence for these claims, I present results of similar experiments I conducted on a public breast MRI dataset, where I demonstrate that the suspected data leakage mechanism closely reproduces the results reported on in the cited work.
- 30 April 2024
- Published by Springer Nature in Journal of Imaging Informatics in Medicine
- Vol. 37 (5) , 2706-2708
Letter to the Editor Regarding Article “Prior to Initiation of Chemotherapy, Can We Predict Breast Tumor Response? Deep Learning Convolutional Neural Networks Approach Using a Breast MRI Tumor Dataset”
The cited article reports on a convolutional neural network trained to predict response to neoadjuvant chemotherapy from pre-treatment breast MRI scans. The proposed algorithm attains impressive performance on the test dataset with a mean Area Under the Receiver-Operating Characteristic curve of 0.98 and a mean accuracy of 88%. In this letter, I raise concerns that the reported results can be explained by inadvertent data leakage between training and test datasets. More precisely, I conjecture that the random split of the full dataset in training and test sets did not occur on a patient level, but rather on the level of 2D MRI slices. This allows the neural network to “memorize” a patient’s anatomy and their treatment outcome, as opposed to discovering useful features for treatment response prediction. To provide evidence for these claims, I present results of similar experiments I conducted on a public breast MRI dataset, where I demonstrate that the suspected data leakage mechanism closely reproduces the results reported on in the cited work.
- 5 April 2024
- Published by Wiley in Journal of Magnetic Resonance Imaging
- Vol. 60 (6) , 2290-2308
AI Applications to Breast MRI: Today and Tomorrow
In breast imaging, there is an unrelenting increase in the demand for breast imaging services, partly explained by continuous expanding imaging indications in breast diagnosis and treatment. As the human workforce providing these services is not growing at the same rate, the implementation of artificial intelligence (AI) in breast imaging has gained significant momentum to maximize workflow efficiency and increase productivity while concurrently improving diagnostic accuracy and patient outcomes. Thus far, the implementation of AI in breast imaging is at the most advanced stage with mammography and digital breast tomosynthesis techniques, followed by ultrasound, whereas the implementation of AI in breast magnetic resonance imaging (MRI) is not moving along as rapidly due to the complexity of MRI examinations and fewer available dataset. Nevertheless, there is persisting interest in AI‐enhanced breast MRI applications, even as the use of and indications of breast MRI continue to expand. This review presents an overview of the basic concepts of AI imaging analysis and subsequently reviews the use cases for AI‐enhanced MRI interpretation, that is, breast MRI triaging and lesion detection, lesion classification, prediction of treatment response, risk assessment, and image quality. Finally, it provides an outlook on the barriers and facilitators for the adoption of AI in breast MRI.Level of Evidence: 5Technical Efficacy: Stage 6
- 5 April 2024
- Published by Wiley in Journal of Magnetic Resonance Imaging
- Vol. 60 (6) , 2290-2308
AI Applications to Breast MRI: Today and Tomorrow
In breast imaging, there is an unrelenting increase in the demand for breast imaging services, partly explained by continuous expanding imaging indications in breast diagnosis and treatment. As the human workforce providing these services is not growing at the same rate, the implementation of artificial intelligence (AI) in breast imaging has gained significant momentum to maximize workflow efficiency and increase productivity while concurrently improving diagnostic accuracy and patient outcomes. Thus far, the implementation of AI in breast imaging is at the most advanced stage with mammography and digital breast tomosynthesis techniques, followed by ultrasound, whereas the implementation of AI in breast magnetic resonance imaging (MRI) is not moving along as rapidly due to the complexity of MRI examinations and fewer available dataset. Nevertheless, there is persisting interest in AI‐enhanced breast MRI applications, even as the use of and indications of breast MRI continue to expand. This review presents an overview of the basic concepts of AI imaging analysis and subsequently reviews the use cases for AI‐enhanced MRI interpretation, that is, breast MRI triaging and lesion detection, lesion classification, prediction of treatment response, risk assessment, and image quality. Finally, it provides an outlook on the barriers and facilitators for the adoption of AI in breast MRI.Level of Evidence: 5Technical Efficacy: Stage 6
- Open Access
- 19 March 2024
- Published by Elsevier in Computer Physics Communications
- Vol. 300
Simulating CDT quantum gravity
- 19 March 2024
- Published by Elsevier in Computer Physics Communications
- Vol. 300
Open AccessSimulating CDT quantum gravity
- Open Access
Kandinsky Conformal Prediction: Efficient Calibration of Image Segmentation Algorithms
Image segmentation algorithms can be understood as a collection of pixel classifiers, for which the outcomes of nearby pixels are correlated. Classifier models can be calibrated using Inductive Conformal Prediction, but this requires holding back a sufficiently large calibration dataset for computing the distribution of non-conformity scores of the model's predictions. If one only requires only marginal calibration on the image level, this calibration set consists of all individual pixels in the images available for calibration. However, if the goal is to attain proper calibration for each individual pixel classifier, the calibration set consists of individual images. In a scenario where data are scarce (such as the medical domain), it may not always be possible to set aside sufficiently many images for this pixel-level calibration. The method we propose, dubbed ``Kandinsky calibration'', makes use of the spatial structure present in the distribution of natural images to simultaneously calibrate the classifiers of ``similar'' pixels. This can be seen as an intermediate approach between marginal (imagewise) and conditional (pixelwise) calibration, where non-conformity scores are aggregated over similar image regions, thereby making more efficient use of the images available for calibration. We run experiments on segmentation algorithms trained and calibrated on subsets of the public MS-COCO and Medical Decathlon datasets, demonstrating that Kandinsky calibration method can significantly improve the coverage. When compared to both pixelwise and imagewise calibration on little data, the Kandinsky method achieves much lower coverage errors, indicating the data efficiency of the Kandinsky calibration.
Open AccessKandinsky Conformal Prediction: Efficient Calibration of Image Segmentation Algorithms
Image segmentation algorithms can be understood as a collection of pixel classifiers, for which the outcomes of nearby pixels are correlated. Classifier models can be calibrated using Inductive Conformal Prediction, but this requires holding back a sufficiently large calibration dataset for computing the distribution of non-conformity scores of the model's predictions. If one only requires only marginal calibration on the image level, this calibration set consists of all individual pixels in the images available for calibration. However, if the goal is to attain proper calibration for each individual pixel classifier, the calibration set consists of individual images. In a scenario where data are scarce (such as the medical domain), it may not always be possible to set aside sufficiently many images for this pixel-level calibration. The method we propose, dubbed ``Kandinsky calibration'', makes use of the spatial structure present in the distribution of natural images to simultaneously calibrate the classifiers of ``similar'' pixels. This can be seen as an intermediate approach between marginal (imagewise) and conditional (pixelwise) calibration, where non-conformity scores are aggregated over similar image regions, thereby making more efficient use of the images available for calibration. We run experiments on segmentation algorithms trained and calibrated on subsets of the public MS-COCO and Medical Decathlon datasets, demonstrating that Kandinsky calibration method can significantly improve the coverage. When compared to both pixelwise and imagewise calibration on little data, the Kandinsky method achieves much lower coverage errors, indicating the data efficiency of the Kandinsky calibration.