Abstract
Medical image quality is often compromised by non-uniform illumination and low contrast, hindering both clinical diagnosis and automated analysis. We present StillGAN, an unsupervised framework for medical image enhancement that utilizes a bi-directional Generative Adversarial Network (GAN) constrained by illumination regularization and structural loss. Unlike standard GANs that may distort critical anatomical details, StillGAN preserves structural integrity while normalizing lighting. Evaluated on corneal confocal microscopy, retinal fundus, and endoscopy datasets, StillGAN demonstrates significant improvements in image quality and downstream diagnostic accuracy, outperforming existing unsupervised enhancement methods.
1. Introduction
High-quality medical imaging is essential for accurate diagnosis, yet real-world acquisition often suffers from artifacts like uneven lighting and sensor noise. In large-scale databases, a significant portion of images—up to 12% for fundus and 30% for retinal scans—are deemed unreadable. While unsupervised domain translation models like CycleGAN offer a path to enhancement without paired data, they frequently fail to preserve the fine structural details, such as nerve fibers or microvasculature, that are vital for clinical decision-making. StillGAN addresses this by enforcing local constraints to maintain anatomical topology while correcting global illumination.
2. Methods
StillGAN employs a bi-directional GAN architecture featuring a ResU-Net generator and a PatchGAN discriminator. The ResU-Net architecture combines U-Net's skip connections for detail preservation with residual blocks to facilitate deeper training, while the PatchGAN discriminator evaluates image quality on a patch-by-patch basis to capture high-frequency local details. The framework introduces two novel loss functions to guide the enhancement process:
- Illumination Regularization ($L_{ill}$): This loss minimizes the difference between local patch intensity and global average intensity, effectively creating an illumination error map to ensure uniform lighting across the image.
- Structure Loss ($L_{st}$): Inspired by the Structural Similarity Index (SSIM), this loss ensures the consistency of image topology. By focusing on covariance and standard deviation, it allows for brightness adjustments without altering the underlying anatomical structures (e.g., the branching of a nerve).
3. Experimental Setup and Results
The model was validated across three distinct medical modalities: corneal confocal microscopy (CORN-2), color fundus imaging (iSee), and endoscopy (EASE). The CORN-2 dataset focused on nerve fiber enhancement, while the iSee dataset was used for fovea localization. The EASE dataset, derived from the CVC-EndoSceneStill collection, provided a challenge with its characteristic specular highlights and dark shadows.
StillGAN demonstrated superior performance compared to traditional methods like CLAHE and other deep learning models like EnlightenGAN. In corneal nerve enhancement, the Signal-to-Noise Ratio (SNR) improved by up to 3.76 dB, and nerve segmentation AUC reached 0.922. For retinal fundus images, fovea localization error ($d$) was reduced from 129.00 to 62.87, significantly improving landmark visibility. Most importantly, in a diabetic retinopathy diagnostic task, ophthalmologist accuracy improved from 75.0% on original images to 81.5% when using StillGAN-enhanced images.
4. Conclusion and Future Work
StillGAN provides a robust unsupervised solution for medical image enhancement by prioritizing structural preservation and illumination uniformity. While effective, the model's performance is sensitive to patch-size hyperparameters; if the patch is too small, contrast is lost, and if too large, lighting remains uneven. There is also a risk of color artifacts, such as artificial reddening in endoscopy images. Future research will focus on integrating explicit color consistency constraints and evaluating the model's robustness across diverse pathological conditions in clinical settings.
Summary updated