E-POSTER DETAIL | 11th Asia Pacific Tele Ophthalmology Society Symposium

Title
Deep Learning based Cataract Detection at doorstep using Smart Phone Images using portable device in Rural India

Authors
SWATI TOMAR, ABHAY BHAMAIKAR

Presenting
SWATI TOMAR

PURPOSE:
To develop an efficient and effective deep learning model to detect and classify cataract using a customised dataset captured through smart phone and portable device.

METHODS:
Custom dataset of 400 images was collected using smart phone and interventional portable device. The dataset was collected from a tribal area.The annotations were made by Step Two: The custom dataset of 400 images was categorised into two groups: Cataract and No Cataract. The annotations were made by three ophthalmologists from private hospital independently .The dataset was split into training set of 80 Images and testing set of 20 Images for ‘No Cataract’, and training set of 320 Images and testing set of 80 Images for ‘Cataract’. Training Dataset was trained on various pre-trained Deep Learning models and tested.

METHODS:

Custom dataset of 400 images was collected using smart phone and interventional portable device. The dataset was collected from a tribal area.The annotations were made by Step Two: The custom dataset of 400 images was categorised into two groups: Cataract and No Cataract. The annotations were made by three ophthalmologists from private hospital independently .The dataset was split into training set of 80 Images and testing set of 20 Images for ‘No Cataract’, and training set of 320 Images and testing set of 80 Images for ‘Cataract’. Training Dataset was trained on various pre-trained Deep Learning models and tested.

RESULTS:
Based on the evaluation of the captured images for cataract using smart phone and interventional portable device with Cataract and No Cataract eye images on various pre- trained Deep Learning models for cataract identification, DenseNet169 (14.1 million) achieved the highest accuracy of 80% on binary classification of cataract with MobileNetV2 (3.5 million) and ResNet152 (60.4 million) giving better results with accuracy of 78%.

RESULTS:

Based on the evaluation of the captured images for cataract using smart phone and interventional portable device with Cataract and No Cataract eye images on various pre- trained Deep Learning models for cataract identification, DenseNet169 (14.1 million) achieved the highest accuracy of 80% on binary classification of cataract with MobileNetV2 (3.5 million) and ResNet152 (60.4 million) giving better results with accuracy of 78%.

CONCLUSIONS:
Since the study focuses on building lightweight model to identify cataract, by using pre- processing technique on customised dataset, higher accuracy can be achieved through MobileNetV2 (3.5 million) making cataract screening easy, fast and affordable. The solution can detect cataracts with >80% sensitivity and specificity, which is equivalent to the accuracy of a trained ophthalmologist.