Eye diseases detection using deep learning with BAM attention module

Journal Article

2024

Published in: https://doi.org/10.1007/s11042-023-17839-9

With the changing lifestyle, a large population suffers from eye diseases such as glaucoma,
cataract, and diabetic retinopathy. Therefore, timely detection and classification of the disease
are necessary to minimize vision loss, however, it is time taking task and requires various
tests and physicians’ in-depth analysis. Thus, an accurate automated technique, timely
detection, and classification are needed to cope with the aforementioned challenges. Therefore,
this study proposes a technique based on an improved deep learning algorithm i.e.,
SqueezeNet that uses the eye image’ features to detect various diseases such as cataract,
glaucoma, and diabetic retinopathy simultaneously. In our proposed model, we employed
Bottleneck Attention Module (BAM) with SqueezeNet having an additional layer. Our proposed
attention module utilizes two different ways and effectively extracts the most representative
features and drops the image’s background features of eyes which don’t take part
in the detection of diseases. Moreover, the algorithm is a pre-trained network that doesn’t
require a huge training set, therefore, the existing dataset i.e., ODIR, cataract, ORIGA,
and glaucoma datasets have been utilized for the training and testing. Additionally, crossvalidation
has been employed using the cataract dataset to assess the performance of the
proposed model. The squeezed connections with regularization power help to minimize
the overfitting during the training of eye samples training sets. The proposed algorithm is
a novel and effective technique to report the successful implementation for the early detection
and classification of eye disease images. The algorithm achieved 98.9% accuracy over
the testing dataset and 98.1% accuracy over cross-validation. Various experiments have
been performed to confirm that our proposed algorithm performs significantly to detect and
classify eye diseases than existing state-of-the-art.

• Publisher Name: SPRINGER
• Publishing City: Netherlands
• Volume Number: 83
• Magazine \ Newspaper: Multimedia Tools and Applications
• Pages: 59061–59084