Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Dimitrios Bounias has a total of 128 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Privacy Leakage in Federated Learning in Radiology Reports: Comparative Evaluation of Tokenizer and Batch-Size Privacy Risks

    by Santhosh Parampottupadam, Andrés Martínez Mora, Dimitrios Bounias, Sinem Sav, Klaus Maier-Hein, Ralf Floca

    Background: Federated learning (FL) enables multi-institutional model training on clinical text without sharing raw data; however, gradient inversion methods can reconstruct sensitive information from shared model updates. The extent of such privacy leakage in FL applied to radiology reports, and the role of tokenizer design, remains unclear. Objective: This study aimed to quantify gradient-based reconstruction of radiology report text in an FL setting and to compare privacy risk across 3 transformer tokenization strategies in a controlled, tokenizer-aware evaluation. Methods: Six FL clients trained a GPT-2–style transformer (sequence length 32) on 2 public clinical-text corpora comprising 368,751 diagnostic reports, 98,206 discharge summaries, and 1500 MIMIC-CXR (Medical Information Mart for Intensive Care Chest X-Ray) radiology reports. Models were trained using 3 tokenizers (GPT-2, RadBERT, and LLaMA-2) with batch sizes of 64, 128, and 256. An active malicious-server threat model was assumed, and analytic gradient inversion was applied to recover text. Reconstruction fidelity was measured over 5 runs using exact sentence accuracy, sentence-level bilingual evaluation understudy (S-BLEU), and recall-oriented understudy for gisting evaluation (ROUGE-L). Results: Exact sentence reconstruction ranged from 27% to 75% across tokenizers, datasets, and batch sizes. At batch size 64 on the discharge dataset, accuracy was 64.7% (GPT-2), 70% (RadBERT), and 67.5% (LLaMA-2), decreasing to 27.3%, 28.5%, and 27.5% at batch size 256. S-BLEU declined with increasing batch size (eg, discharge reports from 0.69 to 0.31). Reconstruction fidelity did not differ significantly across tokenizers (all but one of 27 comparisons nonsignificant; none significant after Holm correction), and approximately 75% of clinical concepts were represented in the reconstructed text (a corpus-level upper bound), regardless of tokenizer. Batch size was the dominant factor governing leakage. Conclusions: Under a worst-case malicious server that tampers with the shared model and observes unprotected per-client gradients (no secure aggregation or differential privacy), substantial portions of radiology-report text can be reconstructed, with up to approximately 75% of reconstructed 32-token sequences and 75% of clinical concepts (not direct patient identifiers) recovered from FL gradients. In a controlled ablation holding model architecture fixed, tokenizer choice, including domain-specific tokenizers, did not significantly affect leakage under the evaluated conditions, whereas batch size was the primary determinant, and no tokenizer significantly reduced the risk. Tokenizer selection should therefore not be treated as a privacy safeguard in this setting. Safeguards such as secure aggregation and differential privacy should therefore be evaluated as candidate protections for FL deployments that must satisfy Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR) requirements in radiology natural language processing (NLP); legal compliance additionally depends on organizational safeguards, risk assessment, and governance beyond the scope of this study.

  • Abstract: The Missing Piece

    by Katharina Eckstein, Constantin Ulrich, Michael Baumgartner, Jessica Kächele, Dimitrios Bounias, Tassilo Wald, Ralf Floca, Klaus H. Maier-Hein

  • Open Access

    Privacy Leakage in Federated Learning in Radiology Reports: A Comparative Evaluation of Tokenizer-Driven Privacy Risks (Preprint)

    by Santhosh Parampottupadam, Andrés Martínez, Dimitrios Bounias, Sinem Sav, Klaus Maier-Hein, Ralf Floca

    BACKGROUND Federated learning (FL) enables multi-institutional model training on clinical text without sharing raw data; however, gradient inversion methods can reconstruct sensitive information from shared model updates. The extent of such privacy leakage in FL applied to radiology reports, and the role of tokenizer design, remains unclear. OBJECTIVE To quantify gradient-based reconstruction of radiology report text in an FL setting and to compare privacy risk across three transformer tokenization strategies in a controlled, tokenizer-aware evaluation. METHODS Six FL clients trained a GPT-2–style transformer (117M parameters; sequence length 32) on two public radiology corpora comprising 368,751 diagnostic reports, 98,206 discharge summaries, and 1,500 MIMIC-CXR free-text reports. Models were trained using three tokenizers (GPT-2, RadBERT, LLaMA-2) with batch sizes of 64, 128, and 256. A curious-server threat model was assumed, and analytic gradient inversion was applied to recover text. Reconstruction fidelity was measured over five runs using exact sentence accuracy, S-BLEU, and ROUGE-L. RESULTS Exact sentence reconstruction ranged from 33% to 42% across tokenizers. At batch size 64, accuracy was 42.1% (GPT-2), 42.3% (RadBERT), and 39.4% (LLaMA-2), decreasing to 37.3%, 37.2%, and 34.3% at batch size 256. S-BLEU scores declined with increasing batch size (e.g., GPT-2: 0.44→0.33; RadBERT: 0.48→0.35; LLaMA-2: 0.39→0.30). RadBERT yielded higher reconstruction fidelity and greater recovery of clinical terms, but no tokenizer prevented leakage. CONCLUSIONS Substantial portions of radiology report text can be reconstructed from FL gradients even with larger batch sizes and domain-specific tokenizers. Tokenizer design influences leakage severity and should be incorporated into privacy evaluations for clinical language models. Integrating safeguards such as secure aggregation and differential privacy is necessary to meet HIPAA and GDPR requirements when deploying FL for radiology NLP. CLINICALTRIAL Not applicable.

  • Open Access

    Risk-adjusted training and evaluation for breast cancer detection

    by Dimitrios Bounias, Michael Baumgartner, Peter Neher, Balint Kovacs, Ralf Floca, Lorenz A. Kapsner, Jessica Eberle, Dominique Hadler, Frederik Laun, Sabine Ohlmeyer, et al.


  • Including AI in diffusion-weighted breast MRI has potential to increase reader confidence and reduce workload

    by Dimitrios Bounias, Lina Simons, Michael Baumgartner, Chris Ehring, Peter Neher, Lorenz A Kapsner, Balint Kovacs, Ralf Floca, Paul F Jaeger, Jessica Eberle, et al.

    Objectives: Breast diffusion-weighted imaging (DWI) has shown potential as a standalone imaging technique for certain indications, eg, supplemental screening of women with dense breasts. This study evaluates an artificial intelligence (AI)-powered computer-aided diagnosis (CAD) system for clinical interpretation and workload reduction in breast DWI. Materials and Methods: This retrospective IRB-approved study included: n = 824 examinations for model development (2017-2020) and n = 235 for evaluation (01/2021-06/2021). Readings were performed by three readers using either the AI-CAD or manual readings. BI-RADS-like (Breast Imaging Reporting and Data System) classification was based on DWI. Histopathology served as ground truth. The model was nnDetection-based, trained using 5-fold cross-validation and ensembling. Statistical significance was determined using McNemar’s test. Inter-rater agreement was calculated using Cohen’s kappa. Model performance was calculated using the area under the receiver operating curve (AUC). Results: The AI-augmented approach significantly reduced BI-RADS-like 3 calls in breast DWI by 29% (P =.019) and increased interrater agreement (0.57 ± 0.10 vs 0.49 ± 0.11), while preserving diagnostic accuracy. Two of the three readers detected more malignant lesions (63/69 vs 59/69 and 64/69 vs 62/69) with the AI-CAD. The AI model achieved an AUC of 0.78 (95% CI: [0.72, 0.85]; P <.001), which increased for women at screening age to 0.82 (95% CI: [0.73, 0.90]; P <.001), indicating a potential for workload reduction of 20.9% at 96% sensitivity. Discussion and Conclusion: Breast DWI might benefit from AI support. In our study, AI showed potential for reduction of BI-RADS-like 3 calls and increase of inter-rater agreement. However, given the limited study size, further research is needed.

  • The Missing Piece: A Case for Pre-training in 3D Medical Object Detection

    by Katharina Eckstein, Constantin Ulrich, Michael Baumgartner, Jessica Kächele, Dimitrios Bounias, Tassilo Wald, Ralf Floca, Klaus H. Maier-Hein


  • Cross-Modality Supervised Prostate Segmentation on CBCT for Adaptive Radiotherapy

    by Balint Kovacs, Goran Stanic, Fabian Weykamp, Florian Ebert, Dimitrios Bounias, Bouchra Tawk, Martin Niklas, Jakob Liermann, Oliver Jäkel, Klaus H. Maier-Hein, et al.


  • Open Access

    The Missing Piece: A Case for Pre-Training in 3D Medical Object Detection

    by Katharina Eckstein, Constantin Ulrich, Michael Baumgartner, Jessica Kächele, Dimitrios Bounias, Tassilo Wald, Ralf Floca, Klaus H. Maier-Hein

    Large-scale pre-training holds the promise to advance 3D medical object detection, a crucial component of accurate computer-aided diagnosis. Yet, it remains underexplored compared to segmentation, where pre-training has already demonstrated significant benefits. Existing pre-training approaches for 3D object detection rely on 2D medical data or natural image pre-training, failing to fully leverage 3D volumetric information. In this work, we present the first systematic study of how existing pre-training methods can be integrated into state-of-the-art detection architectures, covering both CNNs and Transformers. Our results show that pre-training consistently improves detection performance across various tasks and datasets. Notably, reconstruction-based self-supervised pre-training outperforms supervised pre-training, while contrastive pre-training provides no clear benefit for 3D medical object detection. Our code is publicly available at: https://github.com/MIC-DKFZ/nnDetection-finetuning.

  • Open Access

    Leveraging foundation models for content-based image retrieval in radiology

    by Stefan Denner, David Zimmerer, Dimitrios Bounias, Markus Bujotzek, Shuhan Xiao, Raphael Stock, Lisa Kausch, Philipp Schader, Tobias Penzkofer, Paul F. Jäger, et al.


  • Open Access

    AI-Based screening for thoracic aortic aneurysms in routine breast MRI

    by Dimitrios Bounias, Tobit Führes, Luise Brock, Johanna Graber, Lorenz A. Kapsner, Andrzej Liebert, Hannes Schreiter, Jessica Eberle, Dominique Hadler, Dominika Skwierawska, et al.

    Prognosis for thoracic aortic aneurysms is significantly worse for women than men, with a higher mortality rate observed among female patients. The increasing use of magnetic resonance breast imaging (MRI) offers a unique opportunity for simultaneous detection of both breast cancer and thoracic aortic aneurysms. We retrospectively validate a fully-automated artificial neural network (ANN) pipeline on 5057 breast MRI examinations from public (Duke University Hospital/EA1141 trial) and in-house (Erlangen University Hospital) data. The ANN, benchmarked against 3D-ground-truth segmentations, clinical reports, and a multireader panel, demonstrates high technical robustness (dice/clDice 0.88-0.91/0.97-0.99) across different vendors and field strengths. The ANN improves aneurysm detection rates by 3.5-fold compared with routine clinical readings, highlighting its potential to improve early diagnosis and patient outcomes. Notably, a higher odds ratio (OR = 2.29, CI: [0.55,9.61]) for thoracic aortic aneurysms is observed in women with breast cancer or breast cancer history, suggesting potential further benefits from integrated simultaneous assessment for cancer and aortic aneurysms.