Cycle Generative Adversarial Aetwork approach for normalization of Gram-stain images for bacteria detection

The Gram staining method is one of the most effective morphological identification procedures for detecting bacteria from direct smear microscopy. This staining process is inexpensive. It aids in diagnosing bacterial infections quickly as it is used for direct clinical sample specimens such as pus, urine, and sputum. The computer-aided diagnostic system aids the clinician by avoiding tedious manual evaluation procedures. However, images captured often suffer from contrast, illumination, and stain variations due to various camera settings and situations. These differences are due to image acquisition conditions, sample quality, and poor staining procedures. These variations affect the diagnosis process, lowering the image analysis performance of the computer-aided diagnosis system. In this context, the present work proposes a novel color normalization approach based on a Cycle Generative Adversarial Network(GAN). We introduce a novel normalization loss function, L_cycm, which is integrated into our dedicated normalization loss, L_N, within the framework of Cycle GAN(CGAN). The proposed method is compared with the state-of-the-art normalization algorithms qualitatively and quantitatively using the KMC dataset. In addition, the study demonstrates the impact of normalization on the Convolutional Neural Network (CNN) -based segmentation and classification process. Furthermore, a bacteria detection framework is proposed based on the U2Net segmentation model and a CNN classifier. The proposed normalization achieved an SSIM score of 0.93 ± 0.07 and PSNR of 29 ± 3.7. The accuracy of the CNN-based classifier improved by 40 % after normalization.