Heinrich Schulz
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Heinrich Schulz has a total of 23 co-authors. Here are the top co-authors by number of shared publications:
- Axel Saalbach
Philips Research
5 shared publications - Bert Arnrich
Digital Health—Connected Healthcare, Hasso Plattner Institute, University of Potsdam, 14482 Potsdam, Germany
2 shared publications - Bjarne Pfitzner
Hasso Plattner Institute (HPI), Universität Potsdam, Digital Health Cluster, Potsdam, Germany
2 shared publications - Joceline Ziegler
Digital Engineering Faculty, University of Potsdam, 14482 Potsdam, Germany
2 shared publications Clemens Bos
Philips Medical Systems
1 shared publicationDaniel Bystrov
Philips Research Laboratories, Sector Technical Systems
1 shared publication
Latest Publications
- Open Access
- 19 March 2025
- Published by MDPI AG in Diagnostics
- Vol. 15 (6) , 777
Impact of Radiologist Experience on AI Annotation Quality in Chest Radiographs: A Comparative Analysis
Background/Objectives: In the burgeoning field of medical imaging and Artificial Intelligence (AI), high-quality annotations for training AI-models are crucial. However, there are still only a few large datasets, as segmentation is time-consuming, experts have limited time. This study investigates how the experience of radiologists affects the quality of annotations. Methods: We randomly collected 53 anonymized chest radiographs. Fifteen readers with varying levels of expertise annotated the anatomical structures of different complexity, pneumonic opacities and central venous catheters (CVC) as examples of pathologies and foreign material. The readers were divided into three groups of five. The groups consisted of medical students (MS), junior professionals (JP) with less than five years of working experience and senior professionals (SP) with more than five years of experience. Each annotation was compared to a gold standard consisting of a consensus annotation of three senior board-certified radiologists. We calculated the Dice coefficient (DSC) and Hausdorff distance (HD) to evaluate annotation quality. Inter- and intrareader variability and time dependencies were investigated using Intraclass Correlation Coefficient (ICC) and Ordinary Least Squares (OLS). Results: Senior professionals generally showed better performance, while medical students had higher variability in their annotations. Significant differences were noted, especially for complex structures (DSC Pneumonic Opacities as mean [standard deviation]: MS: 0.516 [0.246]; SP: 0.631 [0.211]). However, it should be noted that overall deviation and intraclass variance was higher for these structures even for seniors, highlighting the inherent limitations of conventional radiography. Experience showed a positive relationship with annotation quality for VCS and lung but was not a significant factor for other structures. Conclusions: Experience level significantly impacts annotation quality. Senior radiologists provided higher-quality annotations for complex structures, while less experienced readers could still annotate simpler structures with satisfying accuracy. We suggest a mixed-expertise approach, enabling the highly experienced to utilize their knowledge most effectively. With the increase in numbers of examinations, radiology will rely on AI support tools in the future. Therefore, economizing the process of data acquisition and AI-training; for example, by integrating less experienced radiologists, will help to meet the coming challenges.
- 19 March 2025
- Published by MDPI AG in Diagnostics
- Vol. 15 (6) , 777
Open AccessImpact of Radiologist Experience on AI Annotation Quality in Chest Radiographs: A Comparative Analysis
Background/Objectives: In the burgeoning field of medical imaging and Artificial Intelligence (AI), high-quality annotations for training AI-models are crucial. However, there are still only a few large datasets, as segmentation is time-consuming, experts have limited time. This study investigates how the experience of radiologists affects the quality of annotations. Methods: We randomly collected 53 anonymized chest radiographs. Fifteen readers with varying levels of expertise annotated the anatomical structures of different complexity, pneumonic opacities and central venous catheters (CVC) as examples of pathologies and foreign material. The readers were divided into three groups of five. The groups consisted of medical students (MS), junior professionals (JP) with less than five years of working experience and senior professionals (SP) with more than five years of experience. Each annotation was compared to a gold standard consisting of a consensus annotation of three senior board-certified radiologists. We calculated the Dice coefficient (DSC) and Hausdorff distance (HD) to evaluate annotation quality. Inter- and intrareader variability and time dependencies were investigated using Intraclass Correlation Coefficient (ICC) and Ordinary Least Squares (OLS). Results: Senior professionals generally showed better performance, while medical students had higher variability in their annotations. Significant differences were noted, especially for complex structures (DSC Pneumonic Opacities as mean [standard deviation]: MS: 0.516 [0.246]; SP: 0.631 [0.211]). However, it should be noted that overall deviation and intraclass variance was higher for these structures even for seniors, highlighting the inherent limitations of conventional radiography. Experience showed a positive relationship with annotation quality for VCS and lung but was not a significant factor for other structures. Conclusions: Experience level significantly impacts annotation quality. Senior radiologists provided higher-quality annotations for complex structures, while less experienced readers could still annotate simpler structures with satisfying accuracy. We suggest a mixed-expertise approach, enabling the highly experienced to utilize their knowledge most effectively. With the increase in numbers of examinations, radiology will rely on AI support tools in the future. Therefore, economizing the process of data acquisition and AI-training; for example, by integrating less experienced radiologists, will help to meet the coming challenges.
- 16 September 2022
- Published by Springer Nature
- p. 693-703
Bi-directional Encoding for Explicit Centerline Segmentation by Fully-Convolutional Networks
- 16 September 2022
- Published by Springer Nature
- p. 693-703
Bi-directional Encoding for Explicit Centerline Segmentation by Fully-Convolutional Networks
- Open Access
Defending against Reconstruction Attacks through Differentially Private Federated Learning for Classification of Heterogeneous Chest X-ray Data
Privacy regulations and the physical distribution of heterogeneous data are often primary concerns for the development of deep learning models in a medical context. This paper evaluates the feasibility of differentially private federated learning for chest X-ray classification as a defense against data privacy attacks. To the best of our knowledge, we are the first to directly compare the impact of differentially private training on two different neural network architectures, DenseNet121 and ResNet50. Extending the federated learning environments previously analyzed in terms of privacy, we simulated a heterogeneous and imbalanced federated setting by distributing images from the public CheXpert and Mendeley chest X-ray datasets unevenly among 36 clients. Both non-private baseline models achieved an area under the receiver operating characteristic curve (AUC) of on the binary classification task of detecting the presence of a medical finding. We demonstrate that both model architectures are vulnerable to privacy violation by applying image reconstruction attacks to local model updates from individual clients. The attack was particularly successful during later training stages. To mitigate the risk of a privacy breach, we integrated Rényi differential privacy with a Gaussian noise mechanism into local model training. We evaluate model performance and attack vulnerability for privacy budgets . The DenseNet121 achieved the best utility-privacy trade-off with an AUC of for . Model performance deteriorated slightly for individual clients compared to the non-private baseline. The ResNet50 only reached an AUC of in the same privacy setting. Its performance was inferior to that of the DenseNet121 for all considered privacy constraints, suggesting that the DenseNet121 architecture is more robust to differentially private training.Open AccessDefending against Reconstruction Attacks through Differentially Private Federated Learning for Classification of Heterogeneous Chest X-ray Data
Privacy regulations and the physical distribution of heterogeneous data are often primary concerns for the development of deep learning models in a medical context. This paper evaluates the feasibility of differentially private federated learning for chest X-ray classification as a defense against data privacy attacks. To the best of our knowledge, we are the first to directly compare the impact of differentially private training on two different neural network architectures, DenseNet121 and ResNet50. Extending the federated learning environments previously analyzed in terms of privacy, we simulated a heterogeneous and imbalanced federated setting by distributing images from the public CheXpert and Mendeley chest X-ray datasets unevenly among 36 clients. Both non-private baseline models achieved an area under the receiver operating characteristic curve (AUC) of on the binary classification task of detecting the presence of a medical finding. We demonstrate that both model architectures are vulnerable to privacy violation by applying image reconstruction attacks to local model updates from individual clients. The attack was particularly successful during later training stages. To mitigate the risk of a privacy breach, we integrated Rényi differential privacy with a Gaussian noise mechanism into local model training. We evaluate model performance and attack vulnerability for privacy budgets . The DenseNet121 achieved the best utility-privacy trade-off with an AUC of for . Model performance deteriorated slightly for individual clients compared to the non-private baseline. The ResNet50 only reached an AUC of in the same privacy setting. Its performance was inferior to that of the DenseNet121 for all considered privacy constraints, suggesting that the DenseNet121 architecture is more robust to differentially private training. - Open Access
Defending against Reconstruction Attacks through Differentially Private Federated Learning for Classification of Heterogeneous Chest X-Ray Data
Privacy regulations and the physical distribution of heterogeneous data are often primary concerns for the development of deep learning models in a medical context. This paper evaluates the feasibility of differentially private federated learning for chest X-ray classification as a defense against data privacy attacks. To the best of our knowledge, we are the first to directly compare the impact of differentially private training on two different neural network architectures, DenseNet121 and ResNet50. Extending the federated learning environments previously analyzed in terms of privacy, we simulated a heterogeneous and imbalanced federated setting by distributing images from the public CheXpert and Mendeley chest X-ray datasets unevenly among 36 clients. Both non-private baseline models achieved an area under the receiver operating characteristic curve (AUC) of $0.94$ on the binary classification task of detecting the presence of a medical finding. We demonstrate that both model architectures are vulnerable to privacy violation by applying image reconstruction attacks to local model updates from individual clients. The attack was particularly successful during later training stages. To mitigate the risk of privacy breach, we integrated R\'enyi differential privacy with a Gaussian noise mechanism into local model training. We evaluate model performance and attack vulnerability for privacy budgets $\epsilon \in$ {1, 3, 6, 10}. The DenseNet121 achieved the best utility-privacy trade-off with an AUC of $0.94$ for $\epsilon$ = 6. Model performance deteriorated slightly for individual clients compared to the non-private baseline. The ResNet50 only reached an AUC of $0.76$ in the same privacy setting. Its performance was inferior to that of the DenseNet121 for all considered privacy constraints, suggesting that the DenseNet121 architecture is more robust to differentially private training.Open AccessDefending against Reconstruction Attacks through Differentially Private Federated Learning for Classification of Heterogeneous Chest X-Ray Data
Privacy regulations and the physical distribution of heterogeneous data are often primary concerns for the development of deep learning models in a medical context. This paper evaluates the feasibility of differentially private federated learning for chest X-ray classification as a defense against data privacy attacks. To the best of our knowledge, we are the first to directly compare the impact of differentially private training on two different neural network architectures, DenseNet121 and ResNet50. Extending the federated learning environments previously analyzed in terms of privacy, we simulated a heterogeneous and imbalanced federated setting by distributing images from the public CheXpert and Mendeley chest X-ray datasets unevenly among 36 clients. Both non-private baseline models achieved an area under the receiver operating characteristic curve (AUC) of $0.94$ on the binary classification task of detecting the presence of a medical finding. We demonstrate that both model architectures are vulnerable to privacy violation by applying image reconstruction attacks to local model updates from individual clients. The attack was particularly successful during later training stages. To mitigate the risk of privacy breach, we integrated R\'enyi differential privacy with a Gaussian noise mechanism into local model training. We evaluate model performance and attack vulnerability for privacy budgets $\epsilon \in$ {1, 3, 6, 10}. The DenseNet121 achieved the best utility-privacy trade-off with an AUC of $0.94$ for $\epsilon$ = 6. Model performance deteriorated slightly for individual clients compared to the non-private baseline. The ResNet50 only reached an AUC of $0.76$ in the same privacy setting. Its performance was inferior to that of the DenseNet121 for all considered privacy constraints, suggesting that the DenseNet121 architecture is more robust to differentially private training. - Open Access
Continual Learning for Domain Adaptation in Chest X-ray Classification
Over the last years, Deep Learning has been successfully applied to a broad range of medical applications. Especially in the context of chest X-ray classification, results have been reported which are on par, or even superior to experienced radiologists. Despite this success in controlled experimental environments, it has been noted that the ability of Deep Learning models to generalize to data from a new domain (with potentially different tasks) is often limited. In order to address this challenge, we investigate techniques from the field of Continual Learning (CL) including Joint Training (JT), Elastic Weight Consolidation (EWC) and Learning Without Forgetting (LWF). Using the ChestX-ray14 and the MIMIC-CXR datasets, we demonstrate empirically that these methods provide promising options to improve the performance of Deep Learning models on a target domain and to mitigate effectively catastrophic forgetting for the source domain. To this end, the best overall performance was obtained using JT, while for LWF competitive results could be achieved - even without accessing data from the source domain.
Open AccessContinual Learning for Domain Adaptation in Chest X-ray Classification
Over the last years, Deep Learning has been successfully applied to a broad range of medical applications. Especially in the context of chest X-ray classification, results have been reported which are on par, or even superior to experienced radiologists. Despite this success in controlled experimental environments, it has been noted that the ability of Deep Learning models to generalize to data from a new domain (with potentially different tasks) is often limited. In order to address this challenge, we investigate techniques from the field of Continual Learning (CL) including Joint Training (JT), Elastic Weight Consolidation (EWC) and Learning Without Forgetting (LWF). Using the ChestX-ray14 and the MIMIC-CXR datasets, we demonstrate empirically that these methods provide promising options to improve the performance of Deep Learning models on a target domain and to mitigate effectively catastrophic forgetting for the source domain. To this end, the best overall performance was obtained using JT, while for LWF competitive results could be achieved - even without accessing data from the source domain.
- 29 May 2014
- Published by Wiley in Medical Physics
- Vol. 41 (6Part8) , 188
SU-E-J-141: Comparison of Dose Calculation On Automatically Generated MRBased ED Maps and Corresponding Patient CT for Clinical Prostate EBRT Plans
- 29 May 2014
- Published by Wiley in Medical Physics
- Vol. 41 (6Part8) , 188
SU-E-J-141: Comparison of Dose Calculation On Automatically Generated MRBased ED Maps and Corresponding Patient CT for Clinical Prostate EBRT Plans
- 1 January 2005
- Published by Springer Nature in Lecture Notes in Computer Science
- Vol. 8, 467-473
Real-Time Interactive Viewing of 4D Kinematic MR Joint Studies
Assessment of soft tissue in normal and abnormal joint motion today gets feasible by acquiring time series of 3D MRI images. However, slice-by-slice viewing of such 4D kinematic images is cumbersome, and does not allow appreciating the movement in a convenient way. Simply presenting slice data in a cine-loop will be compromised by through-plane displacements of anatomy and “jerks” between frames, both of which hamper visual analysis of the movement. To overcome these limitations, we have implemented a demonstrator for viewing 4D kinematic MRI datasets. It allows to view any user defined anatomical structure from any viewing perspective in real-time. Smoothly displaying the movement in a cine-loop is realized by image post processing, fixing any user defined anatomical structure after image acquisition.
- 1 January 2005
- Published by Springer Nature in Lecture Notes in Computer Science
- Vol. 8, 467-473
Real-Time Interactive Viewing of 4D Kinematic MR Joint Studies
Assessment of soft tissue in normal and abnormal joint motion today gets feasible by acquiring time series of 3D MRI images. However, slice-by-slice viewing of such 4D kinematic images is cumbersome, and does not allow appreciating the movement in a convenient way. Simply presenting slice data in a cine-loop will be compromised by through-plane displacements of anatomy and “jerks” between frames, both of which hamper visual analysis of the movement. To overcome these limitations, we have implemented a demonstrator for viewing 4D kinematic MRI datasets. It allows to view any user defined anatomical structure from any viewing perspective in real-time. Smoothly displaying the movement in a cine-loop is realized by image post processing, fixing any user defined anatomical structure after image acquisition.