Source-linked AI summary
Generative Adversarial Network for Medical Images (MI-GAN)
Talha Iqbal, Hazrat Ali
TL;DR
Medical-image analysis is constrained by the need for large, expertly annotated datasets, while small datasets can lead to over-fitting. MI-GAN generates realistic retinal images and segmented masks from limited examples, and the authors report improved segmentation performance over existing methods, including Dice coefficients of 0.837 on STARE and 0.832 on DRIVE.
Problem
Medical-image analysis requires costly expert annotations and large datasets, while deep learning can overfit and generalize poorly on small medical-image datasets.
Method
MI-GAN generates realistic retinal images and segmented masks using a GAN with an encoder-decoder architecture and a 400-dimensional noise code.
Results
The method outperformed existing models on AUC ROC, AUC PR, and Dice coefficient, using only 10 training examples from each of DRIVE and STARE.
Takeaways & Limitations
The generated images are visually close to real images, and using them as additional training data enhances image-segmentation performance.
Abstract
from arXiv · showhide
Deep learning algorithms produces state-of-the-art results for different machine learning and computer vision tasks. To perform well on a given task, these algorithms require large dataset for training. However, deep learning algorithms lack generalization and suffer from over-fitting whenever trained on small dataset, especially when one is dealing with medical images. For supervised image analysis in medical imaging, having image data along with their corresponding annotated ground-truths is costly as well as time consuming since annotations of the data is done by medical experts manually. In this paper, we propose a new Generative Adversarial Network for Medical Imaging (MI-GAN). The MI-GAN generates synthetic medical images and their segmented masks, which can then be used for the application of supervised analysis of medical images. Particularly, we present MI-GAN for synthesis of retinal images. The proposed method generates precise segmented images better than the existing techniques. The proposed model achieves a dice coefficient of 0.837 on STARE dataset and 0.832 on DRIVE dataset which is state-of-the-art performance on both the datasets.
1. Introduction
Computer-aided medical imaging and deep learning support retinal vessel analysis, but GAN training can be unstable and medical datasets are limited. MI-GAN addresses these constraints through style transfer, small-sample retinal synthesis, and a modified training schedule.
- Retinal vessel abnormalities can support diagnosis of hypertension, diabetes, RAO, RVO, and other vision-threatening diseases.
- Deep learning methods have shown promising accuracy for automatic medical-image vessel segmentation, sometimes exceeding medical specialists.
- GAN training faces instability because generator–discriminator optimization seeks a Nash equilibrium that determines when training should stop.
- MI-GAN generates realistic retinal images from only tens of examples, unlike a cited method requiring hundreds of training examples.
- The proposed method introduces style transfer based on a particular style representation supplied through an additional input.
- Updating the generator twice as often as the discriminator is intended to accelerate convergence and significantly reduce overall training time.
2. Generative Adversarial Network for Medical Imaging (MI-GAN)
MI-GAN uses a conditional GAN framework to generate retinal images from segmented inputs, combining adversarial, reconstruction, segmentation, and optional style-transfer objectives.
- Generator and discriminator: MI-GAN generates images by conditioning the generator on a segmented binary image y and normally distributed noise z.Varying z is intended to produce plausible, distinct RGB realizations from the same segmented input.
- Generator and discriminator: The discriminator classifies generated images as synthetic and dataset images as real while competing with the generator in alternating optimization.The generator seeks to fool the discriminator, whereas the discriminator separates synthetic from real images.
- Generator and discriminator: A positive trade-off constant λ weights an L1 deviation loss that limits the synthetic image’s deviation from the real image.The generator objective combines adversarial loss with the L1 term.
- Generator and discriminator architecture: The generator uses an encoder-decoder architecture with a 400-dimensional noise code and U-Net skip connections to preserve the morphology of segmented inputs.The architecture uses fixed-size kernels, two strides, no pooling layers, and concatenation through skip connections.
- Segmentation technique: For segmentation, MI-GAN uses gold-standard segmented images and adds binary cross-entropy, combining this segmentation loss with the GAN objective through λ.The added loss penalizes distance between gold-standard and output segmented images.
- Style transfer variant: The style-transfer variant conditions generated images on a target style input while retaining the filamentary content specified by the segmentation image.Its style objective uses VGG-19 features with style, content, and total-variation losses, and replaces λLDEV in the generator objective.
3. Experimental Setup
The evaluation uses DRIVE and STARE retinal-image benchmarks, with standardized resizing and comparisons against existing techniques and discriminator variants.
- Datasets Preparation: DRIVE provides 20 training images at 584 × 565, while STARE provides 10 at 700 × 605.Both datasets are resized to 512 × 512 during preprocessing.
- Datasets Preparation: Figure 4 presents DRIVE images and ground truths above STARE images and ground truths.
- Parameters of proposed model: The generator and discriminator use 4 × 4 × l_f convolutional or de-convolutional filters, with a 400-dimensional noise vector and a 256 × 256 × 64 first layer.
- Parameters of proposed model: Table 1 compares models using different discriminator configurations, while Table 2 compares the proposed method with existing techniques using AUC ROC, AUC PR, and Dice Score.
4. Results and Discussions
MI-GAN is evaluated against discriminator variants and existing retinal-vessel methods, with qualitative comparisons showing close agreement to real images and gold-standard probability structure.
- Quantitative comparison: The proposed method outperformed existing methods on Dice coefficient, AUC ROC, and AUC PR, and surpassed human annotation ability on DRIVE.
- Discriminator comparison: Image GAN outperformed U-Net and Patch GAN, indicating that the strongest discriminator framework produced the best segmentation quality among the compared networks.
- Qualitative comparison: Figure 5 compares fundoscopic images with DRIU and MI-GAN probability maps for DRIVE and STARE.
- Qualitative comparison: MI-GAN generated masks, filamentary structured images, and output images that were visually close to real images.
5. Conclusion
MI-GAN generates realistic retinal vessel images and masks from very small training sets, and using these synthesized images as additional data enhances segmentation performance.
- MI-GAN focuses on retinal vessel image segmentation and generation using synthesized retinal images.
- The framework learns useful features from only 10 training examples per DRIVE and STARE dataset.
- Using synthesized images as additional training data enhances image segmentation performance.
- MI-GAN outperformed existing models on AUC ROC, AUC PR, and Dice coefficient, with fewer false positives on fine vessels and clearer lines.
- Generated output images are visually close to real retinal images.
Compliance with Ethical Standards
The paper reports no declared funding or conflicts of interest. It also states that the work involved neither human participants nor animals, so informed consent was not applicable.
- No funding was declared for the study.
- The authors declared no conflicts of interest.Both Talha Iqbal and Hazrat Ali reported no conflict of interest.
- The article involved no studies with human participants or animals.
- Informed consent was not applicable.