Adversarial and Causal debiasing and adaptation — medical image case-study
Despite the expert-level performance of artificial intelligence (AI) models for various medical imaging tasks, real-world performance failures with disparate outputs for various minority subgroups limit the usefulness of AI in improving patients’ lives. AI has been shown to have a remarkable ability to detect protected attributes of age, sex, and race, while the same models demonstrate bias against historically underserved subgroups of age, sex, and race in disease diagnosis. Therefore, an AI model may take shortcut predictions from these correlations and subsequently generate an outcome that is biased toward certain subgroups even when protected attributes are not explicitly used as inputs into the model. We explore various types of bias from shortcut learning that may occur at different phases of AI model development and develop mitigation technique from preprocessing (data-centric solutions) and during model development (computational solutions) and postprocessing (recalibration of learning).
Domain adaptation – While the kidney segmentation is a required step for various clinical analyses, starting from disease detection, surgical planning, radiation dose optimization, to longitudinal characterization, the model development has largely focused on using a single-phase CT, resulting in significant performance degradation caused by drift in contrast phase appearance. We target three domain shifts to propose a robust kidney segmentation model: (i) contrast to non-contrast, (ii) arterial to venous phase, and (iii) normal functioning kidney to abnormal. Leveraging a traditional segmentation residual network (SegResNet) backbone, we design the domain adaptation for phase resiliency using an auxiliary contrast-phase classifier with confusion loss.
Causal debiasing – Our project introduces MOSCARD, a novel predictive modeling framework that integrates chest X-rays (CXR) and 12-lead electrocardiograms (ECG) to enhance cardiovascular risk assessment. While CXR provides insights into chronic conditions contributing to MACE, ECG captures real-time cardiac electrical activity and structural abnormalities. By combining these two complementary modalities, MOSCARD delivers a more holistic and accurate risk profile than traditional models based on clinical scores, CT scans, or biomarkers. Unlike much of the existing literature, which often overlooks the role of confounders and underlying causal structures in multimodal data, our approach explicitly integrates causal reasoning by incorporating comorbidities and mitigates bias.
Adversarial debiasing – In our ongoing study, we develop a two-step adversarial debiasing approach with partial learning that can reduce the racial disparity while preserving the performance of the targeted task. The methodology has been evaluated onindependent medical image case-studies – skin cancer, chest X-ray, mammograms, and showed promises in bias reduction while preserving the targeted performance.
- Pi, Jialu, Juan Maria Farina, Rimita Lahiri, Jiwoong Jeong, Archana Gurudu, Hyung-Bok Park, Chieh-Ju Chao, Chadi Ayoub, Reza Arsanjani, and Imon Banerjee. “MOSCARD–Causal Reasoning and De-confounding for Multimodal Opportunistic Screening of Cardiovascular Adverse Events.” arXiv preprint arXiv:2506.19174 (2025). MICCAI 2025.
- Pi, Jialu, Juan Maria Farina, Chieh-Ju Chao, Chadi Ayoub, Reza Arsanjani, and Imon Banerjee. “Mitigating Bias in Opportunistic Screening for MACE with Causal Reasoning.” IEEE Transactions on Artificial Intelligence (2025).
- Correa-Medero, Ramon, Umar Ghaffar, Sam Fathizadeh, Bhavik Patel, Haidar Abdul-Muhsin, and Imon Banerjee. “Open-source domain adaptation to handle data shift for volumetric segmentation—use case kidney segmentation.” European Radiology (2025): 1-11.
- Banerjee, Imon, Kamanasish Bhattacharjee, John L. Burns, Hari Trivedi, Saptarshi Purkayastha, Laleh Seyyed-Kalantari, Bhavik N. Patel, Rakesh Shiradkar, and Judy Gichoya. “Shortcuts” causing bias in radiology artificial intelligence: causes, evaluation and mitigation.” Journal of the American College of Radiology (2023).
- Ramasamy, Gokul, Bhavik N. Patel, and Imon Banerjee. “Anomaly Detection using Cascade Variational Autoencoder Coupled with Zero Shot Learning.” In Medical Imaging with Deep Learning, short paper track. 2023.
- Correa-Medero, Ramon L., Bhavik Patel, and Imon Banerjee. “Adversarial Debiasing techniques towards ‘fair’skin lesion classification.” In 2023 11th International IEEE/EMBS Conference on Neural Engineering (NER), pp. 1-4. IEEE, 2023.
- Correra, Ramon, Jiwoong Jason Jeong, Bhavik Patel, Hari Trivedi, Judy W. Gichoya, and Imon Banerjee. “A robust two-step adversarial debiasing with partial learning: medical image case-studies.” In Medical Imaging 2023: Imaging Informatics for Healthcare, Research, and Applications, vol. 12469, pp. 31-38. SPIE, 2023.
- Correa-Medero, Ramón L., Rish Pai, Kingsley Ebare, Daniel D. Buchanan, Mark A. Jenkins, Amanda I. Phipps, Polly A. Newcomb et al. “Causal debiasing for unknown bias in histopathology—A colon cancer use case.” PloS one 19, no. 11 (2024): e0303415.
