Abstract
In telemedicine, the transmission of high-resolution medical images is often constrained by bandwidth. Traditional compression standards like HEVC prioritize visual fidelity across the entire image, often wasting bits on diagnostically irrelevant regions. This paper presents DMIC (Diagnosis-oriented Medical Image Compression), a framework that utilizes hierarchical reinforcement learning and efficient transfer learning to optimize bit-rate allocation for specific diagnostic tasks. By tuning only 2.7% of the model parameters, DMIC achieves a 47.6% bit-rate saving compared to HEVC while maintaining diagnostic accuracy, even when trained on as little as a single annotated image.
Introduction
The rapid growth of remote healthcare has made the efficient transmission of medical data a critical necessity. However, medical images—such as Computed Tomography Angiography (CTA)—require high resolution to ensure diagnostic reliability. Standard compression codecs (HEVC, VVC) are designed for human visual consumption, aiming to minimize pixel-wise distortion. In a clinical context, much of an image may be "diagnostic noise," while specific regions (e.g., coronary arteries) are vital.
Existing task-aware compression models often require massive datasets to train from scratch, which is a significant barrier in medicine due to data privacy and the high cost of expert annotation. DMIC addresses these challenges by leveraging transfer learning from natural video domains to the medical imaging domain, focusing on "semantic" importance rather than just visual reconstruction.
Methodology
The DMIC Framework
DMIC is built upon the Hierarchical Reinforcement Learning based Video Semantic Coding (HRLVSC) framework. It treats the compression process as a sequential decision-making problem, employing two distinct Reinforcement Learning (RL) agents:
- Frame-level Agent: Determines the overall quality and bit budget for an entire frame.
- CTU-level Agent: Operates at the Coding Tree Unit (CTU) level, deciding the specific bit allocation for individual pixel blocks based on their diagnostic relevance.
To guide these agents, the framework uses a downstream diagnostic tool—such as a segmentation algorithm—to generate semantic masks. These masks identify critical regions, allowing the RL agents to learn the optimal relationship between the Quantization Parameter (QP) and diagnostic performance.
Efficient Transfer Learning
A core innovation of DMIC is its approach to domain adaptation. Rather than retraining the entire network on medical data, the authors evaluated seven different tuning strategies. They discovered that tuning only the Advantage Actor-Critic (A2C) module—representing only 28,784 parameters or 2.7% of the total model—yielded superior results. This strategy prevents "catastrophic forgetting," where a model loses the robust feature extraction capabilities learned from large-scale natural video datasets when exposed to small, specialized medical datasets.
Experimental Results
Dataset and Setup
The framework was evaluated using the ImageCAS dataset, a benchmark for coronary artery segmentation. To align with the video-based RL framework, 3D CT slices were processed as sequential video frames.
Performance Metrics
- Bit-rate Savings: DMIC achieved a 47.594% BD-Rate saving compared to the HEVC anchor (HM16.24). This indicates that DMIC can maintain the same level of diagnostic accuracy as standard HEVC while using nearly half the data.
- Few-Shot Robustness: The system demonstrated remarkable efficiency in low-data scenarios. The performance gap between training on a single medical image versus five images was negligible, confirming that the transfer of prior knowledge from natural videos is highly effective for medical tasks.
- Computational Efficiency: The RL-based bit allocation adds a mere 1.52 seconds of overhead for a 33-frame sequence during encoding. Crucially, it adds zero overhead to the decoding process, ensuring that the technology is compatible with existing hardware at the clinician's end.
Discussion and Conclusion
DMIC represents a shift from visual-centric to task-centric compression. By focusing on the diagnostic utility of pixels rather than their aesthetic reconstruction, the framework significantly reduces the bandwidth required for telemedicine. The discovery that tuning a tiny fraction of the model's parameters outperforms full-model tuning provides a viable path for deploying AI in data-scarce medical environments.
Future research will focus on extending DMIC to a broader range of diagnostic tasks, such as tumor detection and fracture analysis, and developing unified models capable of handling multiple diagnostic objectives simultaneously without the need for task-specific retraining.
Summary updated