Mohammed Aljudaibi

Bioinformatics Specialization · Cohort 2024

Enhancing Harmful Covariate Shift Detection in Medical Imaging with LoRA

Detecting harmful covariate shifts (HCS) is crucial for the safe deployment of machine learning models and to ensure their reliability, especially in high-risk domains like healthcare. HCS occurs when the data distribution in deployment differs from the training data, leading to degraded model performance. Existing methods, such as the Detectron, detect HCS with an ensemble of classifiers trained to agree on training data and disagree on a small sample of deployment data. However, this task becomes more challenging with medical imaging data as it would require significant compute and time. By integrating Low-Rank Adaptation (LoRA) into Detectron, we apply LoRA to the base model architecture, allowing each classifier in the ensemble to benefit from the reduced parameter count and improved efficiency. This significantly reduces the size of all of the classifiers in the ensemble of models, making detecting HCS on medical data more feasible and efficient. This method makes the detection of HCS in medical imaging using the Detectron more feasible, and also ensures the robustness of deployed models, providing a reliable mechanism to ensure and maintain model performance in healthcare applications.

See it in the program report (PDF)