Deep learning‐based detection and stage grading for optimising diagnosis of diabetic retinopathy

Abstract

Aims To establish an automated method for identifying referable diabetic retinopathy (DR), defined as moderate nonproliferative DR and above, using deep learning‐based lesion detection and stage grading. Materials and Methods A set of 12,252 eligible fundus images of diabetic patients were manually annotated by 45 licenced ophthalmologists and were randomly split into training, validation, and internal test sets (ratio of 7:1:2). Another set of 565 eligible consecutive clinical fundus images was established as an external test set. For automated referable DR identification, four deep learning models were programmed based on whether two factors were included: DR‐related lesions and DR stages. Sensitivity, specificity and the area under the receiver operating characteristic curve (AUC) were reported for referable DR identification, while precision and recall were reported for lesion detection. Results Adding lesion information to the five‐stage grading model improved the AUC (0.943 vs. 0.938), sensitivity (90.6% vs. 90.5%) and specificity (80.7% vs. 78.5%) of the model for identifying referable DR in the internal test set. Adding stage information to the lesion‐based model increased the AUC (0.943 vs. 0.936) and sensitivity (90.6% vs. 76.7%) of the model for identifying referable DR in the internal test set. Similar trends were also seen in the external test set. DR lesion types with high precision results were preretinal haemorrhage, hard exudate, vitreous haemorrhage, neovascularisation, cotton wool spots and fibrous proliferation. Conclusions The herein described automated model employed DR lesions and stage information to identify referable DR and displayed better diagnostic value than models built without this information.

Automated Detection and Diagnosis of Diabetic Retinopathy: A Comprehensive Survey

“…They reported SE values of 60.7%, 49.5%, 28.3%, 36.3%, 57.3%, 8.7%, 79.8%, and 0.164 over PHE, Ex, VHE, NV, CWS, FIP, IHE, MA, respectively. Quellec et al [ 81 ] focused on four lesions CWS, Ex, HE, and MA using a predefined DCNN architecture named o-O solution and reported the values of 62.4%, 52.2%, 44.9%, and 31.6% over CWS, Ex, HE, and MA for SE, respectively, which shows a slightly better performance for CWS and Ex than Wang et al [ 140 ] and considerably better on MA than Wang et al [ 141 ]. On the other hand, Wang et al [ 141 ] performed better in HE detection.…”

“…Quellec et al 2017[81] focused on four lesions CWS, Ex, HE and MA using a predefined DCNN architecture named o-O solution and reported the values of 62.4 %, 52.2 %, 44.9 % and 31.6 % over CWS, Ex, HE and MA for SE respectively which shows slightly better performance for CWS and Ex than Wang et al 2021[139]. Considerably better for MA than Wang et al 2021[139]. On the other hand, Wang et al performed better in HE detection.…”

Using artificial intelligence reading label system in diabetic retinopathy grading training of junior ophthalmology residents and medical students

“…In our previous studies, during the DR reading training of the AI dataset, we calculated the overall kappa scores for doctors of different seniorities. Seventeen attendings and six consultants in the fundus speciality read 20,503 fundus photographs, and the overall kappa scores were 0.67 for attendings and 0.71 for consultants [ 18 , 21 ]. In our training, the overall kappa score was elevated from 0.67 to 0.81, with a higher score than the attendings and consultants, despite the trainees’ lower levels of professional training.…”

“…In recent years, because of the rapid development of artificial intelligence (AI) techniques, AI techniques based on machine learning play a significant role in DR screening, which acquires high sensitivity and specificity through the learning of a large number of fundus photo training data sets [ 10 – 18 ]. But the fundus photo training data sets needed manual annotation by qualified specialists, and the AI reading results also needed to be confirmed by retina experts.…”

A Decision Support System for Diagnosing Diabetes Using Deep Neural Network

“…The deep learning classification technique makes use of the ResNet v2 CNN architecture ( 24 ), which was trained on tiny patches taken from the entire ear endoscopies before being applied to the complete ear images. A total of four deep learning models were trained for autonomous ascribable diabetic retinopathy detection, dependent on whether or not two criteria were included: DR-related lesions and diabetic retinopathy staging ( 25 ). Table 1 presents a summary of selected works.…”