Research and projects

Project 1

A region-of-interest embedded graph neural architecture for gallbladder cancer detection

Objectives

Gallbladder cancer (GBC) is a reasonably competitive disorder, accounting for almost 50% of biliary tract cancers. Diagnosing GBC is challenging because of its asymptomatic nature in early stages and the similarity in imaging capabilities between benign and malignant gallbladder lesions. Ultrasound imaging, a typically used diagnostic tool, often struggles with issues that include low image quality, noise interference, sensor misalignment, shadows, and misleading textures, in addition to complicating reliable diagnosis. The objective of this study is to use a Region of Interest (ROI)-embedded Graph Neural Network Architecture (RGBNet) to create an advanced artificial intelligence model for the early identification of gallbladder cancer (GBC). Data collection and image pre-processing are the first steps in the procedure, followed by the development of an ROI mask. After that, features are taken out of the ROI and utilized to create graphs, which are fed into RGBNet for accurate GBC detection. RGBNet’s use of ROI improves abnormality identification by precisely extracting tumor regions while reducing interference from unimportant regions. In terms of accuracy (93.09%), precision (90.03%), recall (89.77%), specificity (94.73%), and F1-score (91.15%), RGBNet (with ROI) performs better than state-of-theart models such as MobileNetv3, VGG16, ResNet50, InceptionV3, E˙icientNet-B7, RetinaNet, and DenseNet-264 (with only 10 million parameters). Visualization techniques such as Grad-CAM, Guided Grad-CAM, and Guided Backpropagation are used to explain Read more

Objectives

Knee osteoarthritis (KOA) is a prevalent condition characterized by gradual progression. Physicians use the Kellgren and Lawrence (KL) grading system to understand the severity of an individual condition of KOA. This system categorizes the disease from normal to a severe stage. In this study, we aggregated four datasets to generate an extensive dataset comprising 110,232 raw images by applying an augmentation technique called deep convolutional generative adversarial network(DCGAN). Our proposed approach involved developing a modified compact convolutional transformer (CCT) model known as KOA-CCTNet as the foundational model. The test accuracy for the ResNet50, MobileNetv2, DenseNet201, InceptionV3, and VGG16 was 80.77%, 79.98%, 80.23%, 76.89%, and79.58%,respectively. All of them were surpassed by our proposed KOA-CCTNet model, which had a test accuracy of 94.58% while classifying KOA X-ray images. Read more

 Onion cultivation is a vital part of the global agricultural industry, yet it's often threatened by various leaf diseases that can significantly reduce crop yield and quality. These diseases are traditionally identified by visual inspection, a method that is labor-intensive, time-consuming, and subject to human error. For our study, we used several well-known algorithms (VGG16, ResNet50, InceptionV3, InceptionResNetV2, VGG19, MobilenetV2, and the proposed model) to spot four types of onion diseases (Botrytis Blight, Downy Mildew, Purple Blotch, and Stemphylium Leaf Blight). We had a total of 4484 images after augment which we divided for training, testing, and validation. We also used a technique called Gamma correction to improve the images, along with a green fire blue filter applied for improvement of the test accuracy. In the end, compared to InceptionResnetV2 and MobilenetV2, the proposed model has the highest accuracy 96.86%  Read More

 Agricultural productivity and food security are severely threatened by plant diseases, which affect both yield and quality of crops. The advent of machine learning and deep learning has provided a new avenue for automated detection and diagnosis of these diseases. This review paper delivers an extensive examination of the prevalent trends and prospective developments in employing such technologies. It categorizes various techniques according to the data they utilize, ranging from imagery to sensor data, or a mix of both. The paper reviews renowned datasets prevalent in the research area and assesses the efficacy of various approaches on these datasets. It further discusses the challenges and potential areas for future research, such as the scarcity of data, the interpretability of the algorithms, adaptation to new contexts, and the expansion of these technologies. Furthermore, we highlight potential future Read More

Developed a GAN‑based deep learning architecture, training a generator to create synthetic data and a discriminator to distinguish between real and fake data, resulting in realistic and diverse outputs after 250 epochs  Read More

Frog species across the world are vulnerable and in decline, despite the fact that frogs are an integral part of biological systems. Within this enchanting world, we encounter two intriguing groups: non-poisonous and poisonous frogs. This phenomenon prompts the development of an automated computer vision-based frog detection system that can distinguish between deadly and non-poisonous frogs, leading to the development of early treatment methods and a reduction in relative economic loss. In this study, we present a convolutional neural network-based technique for frog detection. The CNN model required numerous epochs to run in order to provide the best result. Read More