glorybagai/Domain-Shift-Aware-Parameter-Efficient-Fine-Tuning-for-Malaria-Detection-in-Low-Resource-Settings

# Domain-Shift-Aware-Parameter-Efficient-Fine-Tuning-for-Malaria-Detection-in-Low-Resource-Settings # Domain-Shift-Aware Parameter-Efficient Fine-Tuning for Malaria Detection in Low-Resource Clinical Settings > **DSA-LoRA** selects LoRA adapter layers using Maximum Mean Discrepancy (MMD) to concentrate fine-tuning capacity on transformer layers most affected by domain shift, enabling efficient adaptation of medical foundation models to tropical disease imaging with as few as 10 labeled LMIC examples. --- ## 1. What is the question you want to answer? When a vision transformer pretrained on high-income country (HIC) medical imaging data is fine-tuned for malaria detection using a small number of labeled examples from a low-and-middle-income country (LMIC) clinical setting, does the placement of LoRA adapters across transformer layers affect adaptation performance — and can Maximum Mean Discrepancy (MMD) be used to identify which layers should be adapted? Specifically: **standard LoRA applies adapters uniformly across all transformer layers regardless of where domain shift actually occurs. Is this uniform placement optimal under extreme domain shift and data scarcity, or does a principled, data-driven layer selection strategy produce better results with fewer adapter parameters?** --- ## 2. Why is this question important? Malaria kills over 600,000 people annually, with the vast majority of deaths occurring in sub-Saharan Africa and other low-resource regions. Automated malaria detection from blood smear microscopy has the potential to reduce diagnostic delays in settings with few trained pathologists. However, AI models developed on clean, well-resourced HIC datasets fail to generalise to LMIC clinical conditions — where imaging equipment is older, image quality is lower, and labeled training data is extremely scarce. Parameter-efficient fine-tuning methods like LoRA offer a path to adapting large pretrained models with minimal data and compute. But no prio …