PRPreity

Preity

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Preity has a total of 4 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • ODC-KAN: Ocular Disease Classification Using Kolmogorov-Arnold Networks From Fundus Images

    by Preity, Ashish Kumar Bhandari, Syed Shahnawazuddin

    Kolmogorov-Arnold Networks (KANs) offers a novel approach that implements learnable functions on network edges instead of node-based activation employed in traditional neural networks. This paper investigates the suitability and effectiveness of KANs for medical image analysis, particularly on the categorization of retinal disorders. We have contrasted the performance of networks with conventional convolutional and linear layers with a variety of network architectures that apply KAN layers. A total of nine model configurations are studied for better statistical analysis. Our results are intended to illustrate the potential of KANs in the classification of retinal diseases, pointing out important directions for future studies. Qualitative as well as quantitative analyses have been done in this work for better understanding. In the area of retinal disease prediction, we have evaluated the performance on binary as well multi-class classification problems. The experiments are validated on four different datasets, namely ACRIMA, Eye-Diseases-Classification, Messidor, and Harvard Dataverse V1 (HDV1). These datasets consist of images corresponding to healthy eyes as well as those affected with 3 different diseases, namely Cataract, Glaucoma and Diabetic Retinopathy (DR). Furthermore, severity grading of Glaucoma and DR has also been studied in this work. Our experimentation on automatic prediction of ocular disease reveals the advantages of using KAN-based architecture over the conventional neural networks.

  • Attention Inspired Network for Optic Nerve Head Segmentation and Glaucoma Assessment

    by Preity, Ashish Kumar Bhandari, Syed Shahnawazuddin

    Glaucoma, a major cause of irreversible blindness, requires early detection for effective treatment. Accurate segmentation of optic disc (OD) and optic cup (OC) from retinal fundus images is essential for glaucoma diagnosis, enabling the calculation of critical parameters. In this work, we have introduced a hybrid deep-learning-based approach that combines a modified residual network and an attention network with the conventional U-Net architecture for segmenting OD and OC. Structural analysis is done for glaucoma prediction based on the segmented regions. In contrast with the classical approaches that focus on CDR calculation, in this work, we first segment the OD as well as OC regions and then decide whether it is a suspicious case of glaucoma or not. For this purpose, we employ the inferior-superior-nasal-temporal (ISNT) rule along with the determination of the disc damage likelihood scale (DDLS). When a suspicious case is detected, we go for CDR calculation. In order to further enhance performance, in this work, we have employed an image enhancement technique along with several data augmentation techniques. Enhancing and expanding the training dataset leads to more robust parameter estimation. The proposed methodology is rigorously tested on four benchmark datasets namely DRISHTI, RIMONE, ORIGA and REFUGE. At the same time, effect of training one corpus and then testing using different dataset is also studied. In each of the cases, structural parameters are calculated for glaucoma prediction. Extensive experiments presented in this work demonstrate that our proposed approach significantly outperforms other state-of-the-art techniques.

  • SA-U-KAN: Spatial Attention Guided Kolmogorov–Arnold Networks for Optic Disc and Cup Segmentation

    by Preity, Ayushi Shukla, Ashish Kumar Bhandari, Syed Shahnawazuddin

    Optic disc and cup are important structures of human eye and the deformities occurring to these two regions lead to an irreversible disease called glaucoma. Accurate segmentation and analysis are one of the methods to diagnose glaucoma. In this letter, we introduce SA-U-KAN, a novel deep learning architecture that combines convolutional feature extractors, spatial attention modules, and Kolmogorov–Arnold networks (KANs) with the U-Net. The encoder stage of the SA-U-KAN comprises convolutional blocks with spatial attention to extract and refine local features. In addition to that, at the bottleneck stage, a KAN-based tokenization mechanism is used to model complex nonlinearities through interpretable univariate function decompositions. Finally, in the decoder stage, segmentation maps are constructed using skip connections along with attention module to preserve multiscale information. By fusing spatial attention and KAN, SAU-KAN is able to effectively capture local textures and global structures. Experimental results demonstrate the superiority of SAU-KAN over existing techniques, yielding improvements of 1.5% in Dice score (DS) and 2% in intersection of union (IoU) on the RIMONE dataset, and 3.5% (DS) and 4.5% (IoU) on the DRISHTI dataset with 6.9G FLOPs.


  • RD-Net: Residual-Dense Network for Glaucoma Prediction Using Structural Features of Optic Nerve Head

    by Preity, Ashish Kumar Bhandari, Akanksha Jha, Syed Shahnawazuddin

    Glaucoma is called as the silent thief of eyesight. It is related to the internal damage of optical nerve head (ONH). For early screening, the simplest way is to analyze the subtle variations in structural features like cup to disk ratio (CDR), disc damage likelihood scale (DDLS), rim width of the inferior, superior, nasal and temporal (ISNT) regions of ONH. This can be done by accurate segmentation of optic disk (OD) and optic cup (OC). In this work, we have introduced a deep learning framework, called Residual Dense network (RD-NET), for disk and cup segmentation. Based on the segmentation results, the structural features are calculated. The proposed design differs from the traditional U-Net in that it utilizes filters with variable sizes and an alternative optimization method throughout the up- and down-sampling processes. The introduced method is a hybrid deep learning model that incorporates dense residual blocks and squeeze excitation block introduced within the conventional U-Net architecture. Unlike the classical approaches that are primarily based on CDR calculation, in this work, we first segment OD and OC using RD-Net and then analyze ISNT and DDLS. Once a suspicious case is detected, we then go for CDR calculation. In addition to developing an efficient segmentation model, six distinct kinds of data augmentation techniques have been also used in this study to increase the amount of training data. This, in turn, leads to a better estimation of model parameters. The model is rigorously trained and tested on four benchmark datasets namely DRISHTI, RIMONE, ORIGA and REFUGE. Subsequently the structural parameters are calculated for glaucoma prediction. The average accuracies are observed to be 0.9940 and 0.9894 for optic disk and cup segmentation, respectively. The extensive experiments presented in this paper show that our method outperforms other existing state-of-the art algorithms.


  • Soft Attention Mechanism Based Network to Extract Blood Vessels From Retinal Image Modality

    by Preity, Ashish Kumar Bhandari, Syed Shahnawazuddin

    In the field of ophthalmology, most of the ocular diseases are associated with vessel abnormalities. Hence, vessel segmentation plays an important role in order to diagnose ocular disease accurately. Deep learning techniques are grabbing attention these days due to their accurate analysis and extraction of retinal vessels. However, the vessels with low contrast, thin vessels, and vessels with improper boundaries are still overlooked. To cover this gap, we have introduced an attention-mechanism-based deep learning network that includes encoder and decoder blocks. Basically, the architecture is an extension of U-net model in which soft attention blocks are introduced in between the down-sampling and up-sampling phases of the network. The attention blocks learn to focus on the target structures and, therefore, the relevant information gets effectively extracted. In our proposed approach, for the preprocessing of the image, we first extracted the green channel and then applied contrast limited adaptive histogram equalization (CLAHE) for contrast enhancement. The preprocessed image was then fed as the input to the model for vessel segmentation. For the quantitative analysis of our model, we have used four benchmark datasets (DRIVE, STARE, CHASE and HRF). A comparative analysis is also done with the state-of-the- art techniques using several performance parameters. The method proposed in this study is found to be more efficient in extracting retinal vessels. The average accuracy achieved by our method is 98%.