Louisa Fay

All Publications

• From Detection to Mitigation: Addressing Bias in Deep Learning Models for Chest X-Ray Diagnosis. Pacific Symposium on Biocomputing. Pacific Symposium on Biocomputing Mottez, C., Fay, L., Varma, M., Ostmeier, S., Langlotz, C. 2026; 31: 538-550

  Abstract

  Deep learning models have shown promise in improving diagnostic accuracy from chest X-rays, but they also risk perpetuating healthcare disparities when performance varies across demographic groups. In this work, we present a comprehensive bias detection and mitigation framework targeting sex, age, and race-based disparities when performing diagnostic tasks with chest X-rays. We extend a recent CNN-XGBoost pipeline to support multi-label classification and evaluate its performance across four medical conditions. We show that replacing the final layer of CNN with an eXtreme Gradient Boosting classifier improves the fairness of the subgroup while maintaining or improving the overall predictive performance. To validate its generalizability, we apply the method to different backbones, namely DenseNet-121 and ResNet-50, and achieve similarly strong performance and fairness outcomes, confirming its model-agnostic design. We further compare this lightweight adapter training method with traditional full-model training bias mitigation techniques, including adversarial training, reweighting, data augmentation, and active learning, and find that our approach offers competitive or superior bias reduction at a fraction of the computational cost. Finally, we show that combining eXtreme Gradient Boosting retraining with active learning yields the largest reduction in bias across all demographic subgroups, both in and out of distribution on the CheXpert and MIMIC datasets, establishing a practical and effective path toward equitable deep learning deployment in clinical radiology.

  View details for DOI 10.1142/9789819824755_0039

  View details for PubMedID 41758167

• MIMM-X: Disentangling Spurious Correlations for Medical Image Analysis Fay, L., Reguigui, H., Yang, B., Gatidis, S., Kuestner, T. edited by Puyol-Anton, E., Ferrante, E., Feragen, A., King, A., Cheplygina, Ganz-Benjaminsen, M., Glocker, B., Petersen, E., Lee, H. SPRINGER INTERNATIONAL PUBLISHING AG. 2026: 94-103

  View details for DOI 10.1007/978-3-032-05870-6_10

  View details for Web of Science ID 001678692600010

• Determinants of ascending aortic morphology: Cross-sectional deep learning-based analysis on 25,073 non-contrast-enhanced NAKO MRI studies. European heart journal. Cardiovascular Imaging Fay, L., Hepp, T., Winkelmann, M. T., Peters, A., Heier, M., Niendorf, T., Pischon, T., Endemann, B., Schulz-Menger, J., Krist, L., Schulze, M. B., Mikolajczyk, R., Wienke, A., Obi, N., Silenou, B. C., Lange, B., Kauczor, H., Lieb, W., Baurecht, H., Leitzmann, M., Trares, K., Brenner, H., Michels, K. B., Jaskulski, S., Volzke, H., Nikolaou, K., Schlett, C. L., Bamberg, F., Lescan, M., Yang, B., Kustner, T., Gatidis, S. 2025

  Abstract

  AIMS: Understanding determinants of thoracic aortic morphology is crucial for precise diagnostics and therapeutic approaches. This study aimed to automatically characterize ascending aortic morphology based on 3D non-contrast-enhanced magnetic resonance angiography (NC-MRA) data from the epidemiological cross-sectional German National Cohort (NAKO) and to investigate possible determinants of mid-ascending aortic diameter (mid-AAoD).METHODS AND RESULTS: Deep learning (DL) automatically segmented the thoracic aorta and ascending aortic length, volume, and diameter was extracted from 25,073 NC-MRAs. Statistical analyses investigated relationships between mid-AAoD and demographic factors, hypertension, diabetes, alcohol, and tobacco consumption. Males exhibited significantly larger mid-AAoD than females (M:35.5±4.8mm, F:33.3±4.5mm). Age and body surface area (BSA) were positively correlated with mid-AAoD (age: male: r=0.20, p<0.001, female: r=0.16, p<0.001; BSA: male: r=0.08, p<0.001, female: r=0.05, p<0.001). Hypertensive and diabetic subjects showed higher mid-AAoD (DeltaHypertension = 2.9 ± 0.5mm; DeltaDiabetes = 1.5 ± 0.6mm). Hypertension was linked to higher mid-AAoD regardless of age and BSA, while diabetes and mid-AAoD were uncorrelated across age-stratified subgroups. Daily alcohol consumption (male: 37.4±5.1mm, female: 35.0±4.8mm) and smoking history exceeding 16.5 pack-years (male: 36.6±5.0mm, female: 33.9±4.3mm) exhibited highest mid-AAoD. Causal analysis (Peter-Clark algorithm) suggested that age, BSA, hypertension, and alcohol consumption are possibly causally related to mid-AAoD, while diabetes and smoking are likely spuriously correlated.CONCLUSIONS: This study demonstrates the potential of DL and causal analysis for understanding ascending aortic morphology. By disentangling observed correlations using causal analysis, this approach identifies possible causal determinants, such as age, BSA, hypertension, and alcohol consumption. These findings can inform targeted diagnostics and preventive strategies, supporting clinical decision-making for cardiovascular health.

  View details for DOI 10.1093/ehjci/jeaf081

  View details for PubMedID 40052574

• Benchmarking Dependence Measures to Prevent Shortcut Learning in Medical Imaging Mueller, S., Fay, L., Koch, L. M., Gatidis, S., Kuestner, T., Berens, P. edited by Xu, Cui, Z., Rekik, Ouyang, Sun, K. SPRINGER INTERNATIONAL PUBLISHING AG. 2025: 53-62

  View details for DOI 10.1007/978-3-031-73290-4_6

  View details for Web of Science ID 001424559300006

• Structuring Radiology Reports: Challenging LLMs with Lightweight Models Moll, J., Fay, L., Azhar, A., Ostmeier, S., Gatidis, S., Lueth, T., Langlotz, C. P., Delbrouck, J. edited by Christodoulopoulos, C., Chakraborty, T., Rose, C., Peng ASSOC COMPUTATIONAL LINGUISTICS-ACL. 2025: 7707-7724

  View details for Web of Science ID 001805531300392