MI-GAN — Generative Adversarial Network for Medical Images
Summary
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. This work presents MI-GAN for synthesis of retinal images. The MI-GAN method generates precise segmented retinal images better than the existing techniques. The MI-GAN 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.
Motivation
Today, majority of the medical professionals use computer-aided medical images for diagnosis purposes. Retinal vessel network analysis gives us information about the status of general system and conditions of the eyes. Ophthalmologists can diagnose early sign of vascular burden due to hypertension and diabetes as well as vision threatening retinal diseases like Retinal Artery Occlusion (RAO) and Retinal Vein Occlusion (RVO) from abnormality in vascular structure. To aid this kind of analysis, automatic vessels segmentation methods have been extensively studied. Recently, deep learning methods have shown potential to produce promising results with higher accuracy, occasionally better than medical specialist in the field of medical imaging. Deep learning also improves efficiency of analyzing data due to its computational and automated nature but most of the medical images are often 3 dimensional (e.g. MRI and CT) and it is difficult as well as inefficient to produce manually annotated images. In general, medical images are inadequate, expensive and offer restricted use due to legal issues (patient privacy). Moreover, the datasets of medical images available publicly often lack consistency in size and annotation. This makes them less useful for training of neural networks, which are data-hungry. This directly limits the development of medical diagnosis systems. So, generation of synthetic images along with their segmented images will help in medical image analysis and provide better diagnosis systems.
Remarks
These synthesized images are realistic looking. When used as additional training dataset, the framework helps to enhance the image segmentation performance. The MI-GAN model is capable of learning useful features from a small training set. In our case the training set consisted of only 10 examples from each dataset namely DRIVE and STARE. MI-GAN had less false positive rate at fine vessels and have drawn more clearer lines, as compared to other methods.
Read full paper at:
or
