Enhanced Generative Adversarial Network for Medical Image Inpainting using Automatic Multi-Task Learning Approach

Disease categorization, severity confirmation, and future hazard avoidance based on patient medical visual data require an automated Computer Aided Diagnosis (CAD) system. However, distorted medical images can greatly damage diagnosis. Thus, medical image classification requires enhancing diagnostic imaging accuracy and reconstructing damaged regions. These issues have lately received attention in the context of medical image inpainting using Deep Learning (DL) methods. The recent Enhanced Generative Adversarial Networks (EGANs) have proven effective image inpainting solutions, but suffered from the complex and non-optimized training process due to mode collapse and non-convergence. In this paper, Optimized Multi-Tasking GAN (OMT-GAN) automatic medical image inpainting model for performance enhancement is proposed. As the name suggests, the OMT-GAN consisted of three individual tasks for the medial image inpainting such as image edge detection, image completion, and organ boundary generation. A lightweight OMT-GAN model that consists of optimized GAN models is designed in this paper. The conventional GAN models optimized in this paper reduce the complexity and enhance the performances by altering the encoder and decoder layers. Training and evaluation of the OMT-GAN model is done on publicly available medical image. Compared to state-of-the-art approaches, the proposed OMT-GAN had superior Structural Similarity Index Measure (SSIM), Peak Signal Noise Ratio (PSNR), Mean Square Error (MSE) and Universal image Quality Index (UQI). The SSIM, PSNR, and UQI is improved by 5.31 %, 4.72 %, and 4.67 % compared to existing methods.