Predicting Renal Disease and Associated Complications Through Deep Learning Using Retinal Fundus Images Linked to Clinical Data

Performance of deep learning for detection of chronic kidney disease from retinal fundus photographs: A systematic review and meta-analysis

“…Finally, 4 records were included in this review based on the full-text assessment. 8,9,15,16 Table 1 summarized the characteristics of the included studies, which enrolled 114,860 subjects in total. All studies divided the study population into two categories (CKD and normal).…”

“…Researchers have different preferences in this regard. Sabanayagam et al 8 selected model with sensitivity and specificity more than 80%, whereas Zhang et al 9 prioritized larger positive and negative predictive values, Kang et al 15 concentrated on early renal function impairment, whereas James et al 16 limited the number of individuals per subgroup, which made a significant difference in the models they developed. Deep learning based on retinal fundus photographs is still a long way from clinical applications.…”

Examination of alternative eGFR definitions on the performance of deep learning models for detection of chronic kidney disease from fundus photographs

“…The recent successes of artificial intelligence (AI) and deep learning (DL) have encouraged the development of alternative methods to screen for CKD. In light of the proposed linkage between the retina and the kidneys [5][6][7], along with the plethora of useful biological information researchers have extracted from the fundus using DL methods [8][9][10], research groups have applied DL to either directly diagnose CKD from fundus photographs [11][12][13][14][15], predict physiological markers from fundus photographs [16], or to create synthetic markers [17] indicative of CKD from fundus images.…”

Examination of alternative eGFR definitions on the performance of deep learning models for detection of chronic kidney disease from fundus photographs

“…The recent successes of arti cial intelligence (AI) and automated analytic techniques such as machine learning (ML) and deep learning (DL) have encouraged the development of alternative methods to screen for CKD. In light of the proposed linkage between the retina and the kidneys 5-7 , along with the plethora of information researchers have extracted from the fundus using DL 8-10 , research groups have applied DL to either directly identify CKD from fundus photographs [11][12][13][14][15][16] , or to derive physiological 17 or synthetic markers 18 which be indicative of CKD from the fundus.…”

“…Although GFR can be measured by methods such as the urinary clearance of inulin 2 , the inherent di culties in this procedure mean that in practice, clinicians rely on inferential methods which use surrogate markers 3 such as estimated glomerular ltration rate (eGFR), in-urine albumin-creatinine ratio, and the underlying physiological context of the symptoms 4 to assess underlying renal functionality.The recent successes of arti cial intelligence (AI) and automated analytic techniques such as machine learning (ML) and deep learning (DL) have encouraged the development of alternative methods to screen for CKD. In light of the proposed linkage between the retina and the kidneys 5-7 , along with the plethora of information researchers have extracted from the fundus using DL 8-10 , research groups have applied DL to either directly identify CKD from fundus photographs [11][12][13][14][15][16] , or to derive physiological 17 or synthetic markers 18 which be indicative of CKD from the fundus.To date, all publications investigating the direct classi cation of CKD from fundus photographs [11][12][13] have used so-called black box, all-in-one classi ers.Despite differences in precise implementation, the unifying theme of this class of models lies in the packaged feature detector and predictor, whose coe cients are updated together simultaneously to reach an optimization target during the training process. Convolutional neural networks (CNN) 19 are the architectures of choice for feature detectors, and multilayer perceptrons (MLP) or single-layer fully connected layers commonly act as predictors.…”