Detection of CSR from Blue Wave Fundus Autofluorescence Images using Deep Neural Network Based on Transfer Learning

Abstract

Fluid clot below the retinal surface is the root cause of Central Serous Retinopathy (CSR), often referred to as Central Serous Chorioretinopathy (CSC). Delicate tissues that absorb sunlight and enable the brain to recognize images make up the retina. This important organ is vulnerable to damage, which could result in blindness and vision loss for the affected person. Therefore, complete visual loss may be reversed and, in some circumstances, may return to normal with early diagnosis discovery. Therefore, timely and precise CSR detection prevents serious damage to the macula and serves as a foundation for the detection of other retinal disorders. Although CSR has been detected using Blue Wave Fundus Autofluorescence (BWFA) images, developing an accurate and efficient computational system is still difficult. This paper focuses on the use of trained Convolutional Neural Networks (CNN) to implement a framework for accurate and automatic CSR recognition from BWFA images. Transfer Learning has been used in conjunction with pre-trained network architectures (VGG19) for classification. Statistical parameter evaluation has been used to investigate the effectiveness of DCNN. For VGG19, the statistic parameters evaluation revealed a classification accuracy of 97.30%, a precision of 99.56%, an F1 score of 97.25%, and a recall of 95.04% when using a BWFA image dataset collected from a local eye hospital in Cochin, Kerala, India. Identification of CSR from BWFA images is not done before. This paper illustrates how the proposed framework might be applied in clinical situations to assist physicians and clinicians in the identification of retinal diseases.