The thesis presents Hybrid deep learning models for Diabetic Retinopathy Grading. The study leverages various
deep learning techniques along with multiresolution analysis for grading of Diabetic Retinopathy. Attempt has
been made to solve various issues in DR grading. In this regard, four contributions are made using Fundus
images. The first three contributions address the DR grading issues while the fourth contribution focuses on how
Generative AI such as GAN can be used for generating samples for DR grading.
1. The first contribution proposes a weighted ensemble model of pre-trained models (Inceptionv3, ResNet50V2
and InceptionResnetV2) combined with soft attention for DR grading. The suggested approach applies transfer
learning to extract features from retinal images from three best-performing models, such as Inceptionv3,
ResNet50V2, and InceptionResNetV2. Hence, an ensemble of best-performing pre-trained models is introduced
to combine the strengths of these models to capture diverse feature representations. Additionally, a soft attention
mechanism is incorporated to guide the model&rsquos focus toward the most informative retinal regions, improving
sensitivity to subtle lesions.
2. The second contribution proposes a hybrid system integrating the Discrete wavelet transform with
Convolutional neural networks for multi-resolution feature extraction. The multi-resolution decomposition
capability of DWT enables detailed feature extraction by breaking down fundus images into low-frequency and
high-frequency components, which are then analyzed through a CNN. The extracted features are subsequently
fed into fully connected layers for classification. This approach captures both fine and coarse details in fundus
images, allowing CNNS to extract meaningful features at different resolutions.
3. The third contribution focuses on improving the detection of mild-stage DR, which is particularly challenging
due to the size of the microaneurysms. A hybrid feature extraction strategy has been proposed where Local Binary
Pattern is applied to enhance texture features, and CLAHE has been applied to enhance vessel features. Further,
CNNs are employed to process these refined features, which are later combined for final classification. This
two-block CNN approach aims to increase the sensitivity of DR detection, particularly in the early stages.
4. The fourth contribution investigates the impact of GAN-generated images on DR grading and introduces
a custom loss function to handle dataset imbalance. In this regard, an Attention-based Balanced GAN has
been proposed to generate synthetic high-quality fundus images. However, the experiments performed show that
the traditional geometric transformations outperform GAN-based augmentation due to the limited realism of
synthetic images. Additionally, a Performance Aware Weighted Loss (PAWL) function is introduced to mitigate
class imbalance, ensuring equitable learning across all DR severity levels.
The above mentioned contributions are validated on standard publicly available datasets namely, IDRiD,
APTOS, DDR, and EyePACS. Various performance measures including sensitivity, specificity, accuracy etc. are
tested and comparative performance analysis is done with respect to the recently reported competent schemes.