The "INDIAai" (National AI Portal of India) provides weekly comprehensive articles highlighting the research contributions of Indian universities and colleges. 

We aim to offer thorough reporting on the AI research contributions made by a single institution every week. This series focuses on the quality of research contributions from each Indian institution.

Ramrao Adik Institute of Technology (RAIT), founded in 1983, is one of the oldest and most prestigious institutions in Navi Mumbai. It is situated on the lush green campus of D. Y. Patil Deemed to be University and offers modern amenities, experienced faculty, and a welcoming learning atmosphere. The institute adheres to the principle of holistic student development by providing a variety of student-centric events that give students numerous learning opportunities. RAIT is associated with D. Y. Patil Deemed University, recognized by the All India Council for Technical Education (AICTE) and certified by the National Assessment and Accreditation Council with an "A++" rating, paving the path for a bright future.

The institute offers undergraduate programs in Computer Engineering, Electronics and Computer Engineering, Electronics and Telecommunication Engineering, Electrical and Instrumentation Engineering, Information Technology, Computer Science and Business Systems, Artificial Intelligence and Data Science, Computer Science and Engineering in Cyber Security, and Computer Science and Engineering in Artificial Intelligence and Machine Learning. RAIT also provides postgraduate and PhD programs in Electronics Engineering, Computer Engineering, Instrumentation Engineering, Information Technology, and Electronics and Telecommunications Engineering.

Let us explore the top AI research contributions from Ramrao Adik Institute of Technology, D Y Patil Deemed to be University, Nerul, Navi Mumbai.

White blood cell image classification using deep learning with canonical correlation analysis

Authors: A.M. Patil, Mukesh D Patil, Gajanan K Birajdar 

Journal: Innovation and Research in BioMedical Engineering (Elsevier) 

White blood cells are essential in monitoring an individual's health status. Blood disease-related opinions require identifying and characterising a blood sample from the patient. In contemporary methodologies, Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and CNN-RNN model fusion are utilized to improve image content comprehension. Training with big data in medical image analysis has prompted us to identify salient characteristics from sample images from start to finish. Techniques utilizing single-cell patch extraction from blood samples to classify blood cells have produced a high success rate. However, these methods cannot resolve the problems associated with overlapping multiple cells. 

This paper implements the Canonical Correlation Analysis (CCA) technique to tackle this issue. The CCA method considers the consequences of overlapping nuclei by simultaneously extracting, learning, and training regions from multiple nuclei. The classification time is decreased, the dimensions of the input images are compressed, and the network converges more rapidly with more precise weight parameters due to the overlapping of blood cell images. The experimental outcomes assessed through the utilization of a publicly accessible database indicate that the CNN and RNN merging model, when combined with canonical correlation analysis, achieves a greater degree of accuracy in comparison to alternative, cutting-edge blood cell classification methods. 

A survey of fractional calculus applications in artificial neural networks 

Authors: M Joshi, Savita Bhosale, Vishwesh A Vyawahare 

Journal: Artificial Intelligence Review (Springer Nature) 

Deep learning, or artificial neural networks (ANN), is the foundation of machine learning. An ANN's capacity for learning and interpolation renders it an optimal instrument for modelling, control, and many other intricate undertakings. In recent times, fractional calculus (FC), which incorporates integrals and derivatives of any order other than integers, has gained significant attention due to its capacity to model memory-type systems. Numerous attempts have been made to investigate the potential of combining these two disciplines; the most prevalent combination is incorporating fractional derivatives into the learning algorithm.

This article examines the application of fractional calculus to convolutional neural networks, radial basis functions, recurrent neural networks, and backpropagation neural networks, among others. A common abbreviation for these ANNs is fractional-order artificial neural networks (FANNs). A comprehensive examination of the diverse principles encompassing FANNs is provided, encompassing activation functions, fractional derivative-based training algorithms, stability, synchronization, hardware implementations of FANNs, and practical applications of FANNs. Additionally, the research emphasizes the benefit of integrating fractional derivatives with ANN, the influence of fractional derivative order on performance metrics such as mean square error, the training and testing duration for FANN, and the stability and synchronization of FANN. 

COVID-19 Respiratory Sound Signal Detection Using HOS-Based Linear Frequency Cepstral Coefficients and Deep Learning 

Authors: Sandeep B. Sangle, Chandrakant J. Gaikwad 

Journal: Circuits, Systems, and Signal Processing (Springer Nature) 

The COVID-19 virus has emerged as a severe threat to human health. Despite numerous reported variants, the virus continues to undergo mutation, and novel strains occasionally emerge. Early COVID-19 detection is a critical concern contributing to the disease's effective management. This research investigates COVID-19 audio signals emanating from vocalizations, sneezing, and respiration—most scholarly works on this subject use MFCC-based features.

This work presents a range of proposed methods for the detection of COVID-19. The proposed methodologies employ accumulated bispectrum characteristics to discern the unique attributes of COVID-19 within the signals above. Three novel approaches are suggested for the detection of COVID-19. A comprehensive analysis of the methods' efficacy is provided, including a comparison with the most recent advancements in the field. Significant performance improvements are observed in the proposed methodologies across various signals. In this investigation, the CNN and ResNet-50 network models are implemented. 

Design of an Optimized Self-Acclimation Graded Boolean PSO with Back Propagation Model and Cuckoo Search Heuristics for Automatic Prediction of Chronic Kidney Disease 

Authors: Anandita Khade, Amarsinh V. Vidhate, Deepali Vidhate 

Journal: Journal of Mobile Multimedia (River Publishers) 

Back Propagation Neural Networks (BPNNs), a type of Artificial Neural Network (ANN), have found widespread implementation across diverse domains, such as stock market prediction, medical diagnosis, and optical character recognition. Numerous studies have utilized BPNN to develop clinical diagnosis decision-support aids for physicians.

Chronic Kidney Disease (CKD) is one such condition that, in recent decades, has gained significant attention owing to the absence of symptoms during its early stages. This study aims to showcase the efficacy of Artificial Intelligence (AI) algorithms in the timely identification of chronic kidney disease (CKD). 

CSIMH: Design of an Efficient Security-Aware Customized Sidechaining Model via Iterative Meta-Heuristics 

Authors: Nisha Balani, Pallavi Chavan 

Journal: Journal of Applied Security Research (Taylor & Francis) 

Blockchains provide a secure alternative for deploying security and long-term data storage. While blockchains possess an infinite capacity, their quality of service (QoS) efficacy deteriorates once a specific threshold of blocks is reached. Thus, sidechains have evolved into a necessity for decentralizing, securing, and facilitating the operation of systems, thereby enhancing their scalability and QoS performance. Sidechains decrease the amount of time required for data storage and extraction. However, all sidechains of a single blockchain are identical in capacity, size, and security. Consequently, their applicability is restricted to practical scenarios that necessitate dynamic security.

To address this constraint, the authors suggest implementing a meta-heuristic strategy for developing a sidechain generation system responsive to the quantity and quality of data stored on the chain. The model uses a machine-learning methodology to determine the optimal sidechain configuration for various data types. It increases the system's scalability, speed, storage, and memory utilization. Multiple data sets evaluate and compare the proposed model to several cutting-edge sidechain deployments. Various assaults and faulty nodes validate the model's security performance. The proposed model demonstrates a consistent quality of service (QoS) performance across various attack categories, thus confirming its resilience to such attacks. Due to this performance enhancement, the proposed model is now suitable for implementation in high-speed, low-energy, and high-security applications such as IoT, mobile ad-hoc networks, and sensor networks.