NLM DIR Seminar Schedule

Abstract:

Medical image processing aided by artificial intelligence (AI) and machine learning (ML) significantly improves the efficiency and performance of medical diagnosis and decision making. However, the difficulty to access well-annotated medical images becomes one of the main constraints on further improving this technology. A recent study reveals that the seemly high-performance deep neural networks (DNNs) for COVID-19 chest X-Ray image detection are vulnerable from network attacks. The reason behind is that the DNNs are optimized by extremely imbalanced datasets where the COVID-19 images only occupy 5% to 6% of the total image. Another drawback is that the medical image patterns are different from the common image patterns. The available pre-optimized DNNs are usually trained by general-purposed image dataset such as the ImageNet. Though the low-level image patterns can be still captured by the bottom filters of the DNN, the high-level differentiable patterns are unlikely to be effectively combined through the complex architecture of the network due to insufficient training examples. The above potential weakness all contributes to the vulnerability of the current DNNs. Therefore, A new approach to improve both the quantity and diversity of the medical image datasets to improve the accuracy and robustness of DNNs.

Generative adversarial network (GAN) is a DNN framework for data synthetization. It becomes a practical solution for medical image generation when a proper constraint is added to control the GAN generator to produce images belonging to a preferred domain. In this study, we propose an adaptive cycle-consistent adversarial network (Ad Cycle GAN) with pretrained DNN to add extra penalty to GAN architecture as a dynamic criterion to control the synthesized medical images to the desired domain while it still good diversity for significant medical patterns. To evaluate the GAN performance, we respectively use a COVID-19 chest X-ray dataset (2,347 images) and a malaria blood cell dataset (19,578 images) to optimize the new Ad Cycle GAN. The synthesized images are evaluated by classification accuracy, Frechet Inception Distance (FID), and subjective evaluation. The initiative results both indicate that over 99% of the synthesize medical images by Ad Cycle GAN are classified to the desired category, which very low FID scores, which reflects they are homogeneous. In addition, the generated images demonstrate enough diversity and good subjective quality given the complexity of the image patterns.

In conclusion, the new Ad Cycle GAN can accurately generate synthetic medical images to the desired domain compared to the original Cycle GAN. The dynamic criterion provides effective control to the GAN architecture to generate desirable medical images with complex discriminative pattern associate with medical expertise.