Introduction
Medical Image Quality Enhancement (IQE) is a critical preprocessing step designed to improve the clarity of clinical scans, such as removing noise or increasing spatial resolution. Traditionally, these methods have focused on optimizing for human perception, using metrics like Peak Signal-to-Noise Ratio (PSNR). However, as medical workflows increasingly rely on automated diagnostic tools, a "perception-recognition paradox" has emerged: images that appear cleaner to the human eye can sometimes lead to poorer performance in downstream AI recognition tasks. This occurs because enhancement algorithms may inadvertently smooth over subtle, high-frequency features essential for machine learning models.
The Challenge: Gradient Interference
To bridge the gap between human perception and machine utility, researchers have explored "task-driven" models that train enhancement and recognition components simultaneously. While promising, these models often suffer from gradient interference. In a joint training environment, the enhancement task (e.g., denoising) and the recognition task (e.g., lesion segmentation) may have conflicting requirements. For example, a denoising objective might push for global smoothness, while a segmentation objective requires sharp edges and local texture preservation. This "tug-of-war" between gradients often leads to sub-optimal convergence for both tasks.
Methodology: Gradient Promotion (GradProm)
This paper introduces GradProm (Gradient Promotion), a novel training strategy designed to resolve these conflicts by selectively integrating task-specific feedback. The framework partitions the system into a Mainstream Model for image enhancement and an Auxiliary Model for visual recognition.
Conditional Gradient Updates
The core innovation of GradProm lies in how it handles the parameter updates for the enhancement model. During training, the system calculates gradients for both the image processing task ($G_{IP}$) and the visual recognition task ($G_{VR}$). It then measures the cosine similarity ($s$) between these two vectors:
- Alignment ($s \ge 0$): If the gradients are aligned or orthogonal, the recognition task is considered helpful. GradProm uses both gradients to update the enhancement model, allowing the recognition task to "promote" the enhancement process.
- Conflict ($s < 0$): If the gradients conflict, the system ignores the recognition gradient entirely, updating the enhancement model using only its own task-specific gradient ($G_{IP}$).
This strategy is supported by a mathematical proof (Lemma 3.1) ensuring that the model converges to a local minimum without being biased or distracted by the auxiliary task's conflicting requirements.
Experimental Validation
The researchers evaluated GradProm across four diverse medical imaging modalities to demonstrate its generalized applicability:
- ISIC 2018: RGB skin lesion images (Denoising + Diagnosis).
- COVID-CT: Grayscale CT scans (Denoising + COVID-19 Detection).
- Lizard: Histology images (Super-resolution + Segmentation).
- CAMUS: 2D echocardiography (Denoising + Structure Segmentation).
Key Results
GradProm consistently outperformed state-of-the-art (SOTA) methods like ESTR across all benchmarks. On the ISIC 2018 dataset, GradProm achieved a PSNR improvement of over 1.2 dB and an accuracy boost of approximately 2% compared to standard joint training under high-noise conditions. Notably, the performance gap between GradProm and baseline methods widened as noise levels increased, highlighting the method's robustness for low-quality scans.
Furthermore, GradProm demonstrated superior domain generalization. In cross-dataset tests (e.g., training on ISIC and testing on Lizard), the model maintained significantly higher performance than competing methods, suggesting that GradProm learns robust, transferable features rather than simply memorizing dataset-specific noise patterns.
Conclusion and Future Work
GradProm offers a powerful, architecture-agnostic strategy for medical image enhancement. By treating downstream recognition as an auxiliary guide rather than a fixed constraint, it effectively resolves the perception-recognition paradox. While the method introduces no additional overhead during inference, it does rely on the availability of a relevant auxiliary model during the training phase. Future research may extend this "gradient promotion" logic to other multi-objective medical imaging tasks, such as image compression or registration.
Summary updated