AdaLoRA-QAT: Adaptive Low-Rank and Quantization-Aware Segmentation
Summary: arXiv:2604.01167v1 Announce Type: cross
Abstract
Chest X-ray (CXR) segmentation is an important step in computer-aided diagnosis, yet deploying large foundation models in clinical settings remains challenging due to computational constraints. We propose AdaLoRA-QAT, a two-stage fine-tuning framework that combines adaptive low-rank encoder adaptation with full quantization-aware training.
Key Features of AdaLoRA-QAT
- Adaptive Rank Allocation: This feature improves parameter efficiency by dynamically adjusting the rank of the model’s components, allowing for more effective utilization of computational resources.
- Selective Mixed-Precision INT8 Quantization: This process preserves the structural fidelity crucial for clinical reliability while significantly reducing the model’s size and computational requirements.
- High Performance: Evaluated across large-scale CXR datasets, AdaLoRA-QAT achieves a remarkable 95.6% Dice score, matching the performance of full-precision SAM decoder fine-tuning.
- Significant Model Compression: The framework reduces trainable parameters by 16.6 times and yields a model compression factor of 2.24 times, making it a compact solution for medical image segmentation.
Research Findings
The evaluation results indicate that AdaLoRA-QAT provides an effective balance between accuracy, efficiency, and structural trustworthiness. A Wilcoxon signed-rank test confirmed that quantization does not significantly degrade segmentation accuracy, which is crucial for clinical applications.
Conclusion
AdaLoRA-QAT showcases the potential of adaptive low-rank and quantization-aware training in enhancing the deployment of AI models in medical settings. By effectively optimizing performance and minimizing resource consumption, this framework paves the way for the use of compact foundation models in real-world clinical environments.
Access to Code and Pretrained Models
For those interested in exploring the AdaLoRA-QAT framework further, code and pretrained models are available at the following link: AdaLoRA-QAT Resource Page.