PhD Defense: Ethan Dack Lays Groundwork for Clinically Meaningful AI Models

In this thesis, Ethan focused on the development of AI-based diagnostic methods for resource-constrained settings, with particular emphasis on dataset bias and its impact on model performance. To do this, he addressed various medical imaging tasks across diverse modalities.

First, he took advantage of the large volumes of unstructured data on COVID-19 generated during the pandemic and used the fine-grained information in chest CT scans to perform zero-shot multi-label classification via contrastive vision–language learning, which he then evaluated in collaboration with clinical experts for their ability to support clinicians in diagnostics. Building on this evaluation, Ethan further examined the role of self-supervised pretraining by training a masked autoencoder on a curated dataset, with subsequent fine-tuning to improve diagnostic performance.

Lastly, he investigated dataset bias in widely used open-source X-ray datasets, which are frequently reused due to the scarcity of medical imaging data. His investigation revealed that modern models often exploit differences in pixel intensity and texture rather than pathology, which underscores the need for more explainable methods and the development of diverse, rigorously curated open-source medical imaging datasets.

With this thesis, Ethan demonstrated the critical role that self-supervised learning is likely to play in the future of medical image analysis, showcasing its ability to capture meaningful and clinically relevant representations that can be effectively leveraged for diagnostic tasks. His findings also identified important limitations associated with relying exclusively on open-source datasets, emphasizing the importance of careful dataset curation to mitigate noisy inputs and demographic bias.

Congratulations to Dr. Ethan Dack on this incredible achievement and for laying the groundwork for the development of robust, reliable, and clinically meaningful AI models in the medical domain!

Publications:

Dack, Ethan (2025). Developing Robust and Generalizable AI for Diagnostic Support Across Medical Imaging Modalities. MICCAI Emerge Scholars (accepted).

Dack, Ethan; Brigato, Lorenzo; Dedousis Vasilis; Gote-Schniering, Janine; Magnin, Cheryl; Hoppe, Hanno; Exadaktylos, Aristomenis; Funke-Chambour, Manuela; Geiser, Thomas; Christe, Andreas; Ebner, Lukas; Mougiakakou, Stavroula (2025). Unmasking Interstitial Lung Diseases: Leveraging Masked Autoencoders for Diagnosis. In: Bhattarai, B. et al. Data Engineering in Medical Imaging. DEMI 2025. Lecture Notes in Computer Science, vol. 16191. Springer Nature 10.1007/978-3-032-08009-7_16

Dack, Ethan and Chengliang, Dai (2025). Understanding Dataset Bias in Medical Imaging: A Case Study on Chest X-rays. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, October 2025, 573–583. IEEE 10.1109/ICCVW69036.2025.00066

Hamedi, Zahra; Brigato, Lorenzo; Dack, Ethan; Schütz, Mira; Lehmann, Beat; Exadaktylos, Aristomenis; Mougiakakou, Stavroula; Krummrey, Gert (2025). AI-Based Analysis of Abdominal Ultrasound Images to Support Medical Diagnosis in Emergency Departments. Studies in Health Technology and Informatics 325:16–21. IOS Press 10.3233.SHTI250209

Fontanellaz, Matthias; Christe, Andreas; Christodoulidis, Stergios; Dack, Ethan; Roos, Justus; Drakopoulos, Dionysios; Sieroń, Dominik; Peters, Alan Arthur; Geiser, Thomas; Funke-Chambour, Manuela; Heverhagen, Johannes; Hoppe, Hanno; Exadaktylos, Aristomenis; Ebner, Lukas; Mougiakakou, Stavroula (2024). Computer-Aided Diagnosis System for Lung-Fibrosis: From the Effect of Radiomic Features and Multi-Layer-Perceptron Mixers to Pre-Clinical Evaluation. IEEE Access 12(2024):25642–25656. IEEE 10.1109/ACCESS.2024.3350430

Dack, Ethan; Brigato, Lorenzo; McMurray, Matthew; Fontanellaz, Matthias; Frauenfelder, Thomas; Hoppe, Hanno; Exadaktylos, Aristomenis; Geiser, Thomas; Funke-Chambour, Manuela; Christe, Andreas; Ebner, Lukas; Mougiakakou, Stavroula (2023). An Empirical Analysis for Zero-Shot Multi-Label Classification on COVID-19 CT Scans and Uncurated Reports. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, October 2023, 2634–2643. arXiv 10.48550/arXiv.2309.01740

Dack, Ethan; Christe, Andreas; Fontanellaz, Matthias; Brigato, Lorenzo; Heverhagen, Johannes; Peters, Alan Arthur; Huber, Adrian Thomas; Hoppe, Hanno; Mougiakakou, Stavroula; Ebner, Lukas (2023). Artificial Intelligence and Interstitial Lung Disease: Diagnosis and Prognosis. Investigative Radiology 58(8):602–609. Wolters Kluwer Health 10.1097/RLI. 0000000000000974