Guangzhou Diabetic Eye Study: rationale, design, methodology and baseline data

Abstract

Purpose: The incidence and risk factors for diabetic retinopathy (DR) in southern China remains unclear. This project aims to explore the onset and progression of DR and their determinants through a prospective cohort in South China. Methods: The Guangzhou Diabetic Eye Study (GDES) recruited patients with type 2 diabetic patients registered in the community health centres in Guangzhou, China. Patients with history of ocular treatment, severe refractive opacity, or other systemic diseases were excluded. Comprehensive examinations were performed including visual acuity, refraction, ocular biometry, fundus imaging, blood and urine tests. The biological samples were collected for further study. Results: A total of 2305 eligible patients were included in the final analysis. In total, 14.58% of the participants had any DR and 4.25% had vision-threatening DR (VTDR), among which 76 (3.30%), 197 (8.55%), 45 (1.95%) and 17 (0.74%) were classified as mild NPDR, moderate NPDR, severe NPDR and PDR, respectively. There were 93 (4.03%) patients of diabetic macular edema (DME). The presence of any DR was independently associated with a longer duration of DM, higher degree of HbA1C, insulin treatment, higher average arterial pressure, higher concentration of serum creatinine, presence of urinary microalbumin, older age, and lower BMI (all P <0.001). For VTDR, 7 factors were significant: older age, a longer duration of DM, higher concentration of HbA1c, use of insulin, lower BMI, higher concentration of serum creatinine, and high albuminuria (all P<0.05). These factors were also independently associated with DME (all P <0.001). Conclusions: The GDES is the first large-scale prospective cohort study of the diabetic population in southern urban China, which will help to identify newer imaging and genetic biomarkers for DR in this population.

SplitNets: Designing Neural Architectures for Efficient Distributed Computing on Head-Mounted Systems

“…SplitNets training / searching configurations In supernet training, we sample 5 kinds of networks (see Appendix D) and optimize for 360 epochs with batch size 4096 using the standard training receipt from [54]. In resource-constrained searching, we sample 512 candidate networks with 20 generations using evolution algorithm [9,54,55].…”

“…As we mentioned in Section 3, We build the supernet's search spaces mainly based on [9,54] with some extra search spaces from our SA-NAS as shown in Figure 3. Search spaces from standard NAS [9,54] include number of layers, output channel size and expand ratio of each inverted residual block (MB). Following [9,54], we use Swish as activation function.…”

“…Search spaces from standard NAS [9,54] include number of layers, output channel size and expand ratio of each inverted residual block (MB). Following [9,54], we use Swish as activation function. The supernet architecture configuration and search space for single-view SplitNets on ImageNet are summarized in Table 7.…”

“…In supernet training, we jointly train multiple subnetworks sampled from the supernet at each iteration. Various sampling strategies can be used [5,9,54,60]. For example, BigNAS [60] samples three kinds of networks: 1) the largest sub-network ("max"); 2) the smallest sub-network ("min"); 3) several randomly sub-sampled networks.…”

“…Since the largest sub-network usually has better accuracy, it can be used as teacher to supervise other sub-networks via knowledge distillation losses [23]. Some variants of this "sandwich" sampling strategy are also proposed like sampling sub-networks according to their predicted accuracy [54].…”