Severity Classification of Diabetic Retinopathy using ShuffleNet

Volumetric Model Genesis in Medical Domain for the Analysis of Multimodality 2-D/3-D Data Based on the Aggregation of Multilevel Features

“…For example, our final MDF-RN+STB-CN model outperforms the secondbest modality-specific method of Srinivasu et al [20] with average gains of 1.98%, 2.85%, 3.69%, 2.04%, and 1.62% in terms of ACC, F1, PRE, REC, and AUC, respectively. Most of the existing methods [18]- [21] exploit only high-level semantic features to make a classification decision. Our network design mainly leverages multiscale dilated convolutions (DF-blocks) and multilevel feature fusion in a mutually beneficial manner and finally achieves superior classification results.…”

“…They considered a larger dataset (including a total of 10015 images) than that of [16]. In the context of diabetic retinopathy (DR), Gambhir et al [18] proposed a severity classification CAD method that was able to detect and distinguish DR into different severity levels. An existing ShuffleNet [30] model was trained to categorize the input DR image into one of five different classes (including one normal and four diseased cases).…”

“…Most of the existing methods [18]- [21] are disease and modality-specific, and optimized for a limited number of data samples. Moreover, these methods are designed to make diagnostic decisions based on 2D imaging data employing image-based classification models [22]- [31], even in case of 3D imaging data.…”

“…The regularity of this comparative study is defined as follows: Initially, we compared the performance of our first image-based model with various image-based classification methods [7]- [18] using our selected dataset. All these comparative models are labeled as image-based models in Table V.…”

“…With respect to image-and sequence-based CAD methods, convolutional neural networks (CNNs), a wellknown variant of DL algorithms, have received special consideration. Various types of CNNs [15]- [18] have been proposed in the literature for general and medical applications. The structure of a CNN model mainly comprises multiple convolutional layers and fully connected (FC) layers, including trainable parameters [9].…”