Computer Science, Computer Vision and Pattern Recognition, Artificial Intelligence, Information Systems
7
Scopus Publications
Scopus Publications
Explainable Lightweight Transformer-Based Neural Network for Multi-Label Medical Image Classification Chilukamari Rajesh, Chintakindi Balaram Murthy, Medipelly Rampavan, Srinivas Arukonda Transformative Role of Transformer Models in Healthcare, 2025 Accurately classifying medical images with multiple labels is essential for early disease detection and enhancing clinical decision-making. In contrast to single-label classification, multi-label approaches allow for the simultaneous identification of multiple co-existing pathologies in a single image. Deep learning approaches, including convolutional neural networks and transformer-based models, have shown promising results, but they often suffer from high computational costs and lack of explainability, making them impractical for many medical applications. To address these challenges, this study introduces a novel lightweight transformer-based neural network optimized for multi-label medical image classification, reducing computational complexity while preserving strong feature extraction capabilities. Evaluations on the ChestX-ray11 dataset show superior classification accuracy and computational efficiency compared to existing methods. Furthermore, Grad-CAM++ visualizations enhance interpretability by highlighting disease-relevant regions, fostering trust in medical AI applications.
Classification of Cataract Fundus Images Using CataractDQN Chintakindi Balaram Murthy, Chilukamari Rajesh, Kankanala Srinivas, Sathish Mothe Proceedings 2025 International Conference on Emerging Information Technology and Engineering Solutions Eites 2025, 2025 Cataracts pose a significant global health challenge, contributing to a substantial portion of visual impairment and blindness cases worldwide. This paper presents a comprehensive exploration of cataract detection using deep reinforcement learning (DRL) techniques. Leveraging the principles of Deep Q-Network (DQN), the proposed Cataract DQN architecture integrates a convolutional neural network (CNN) comprising nine layers, designed for feature extraction and binary classification of cataract presence. The DQN framework enables iterative refinement of action selection and policy learning through exploration and exploitation strategies, experience replay, and target networks. The model demonstrates promising accuracy rates, reaching up to 95.64% through experimentation, highlighting its potential for enhancing cataract diagnosis and enabling early intervention. Future research avenues include extending the model to classify other eye diseases, addressing grading and localization challenges, and enhancing overall eye care through innovative deep-learning approaches.
Automated Deep Learning Models for Medical Image Segmentation and Denoising Chilukamari Rajesh, Ravichandra Sadam, Sushil Kumar 2024 17th International Conference on Signal Processing and Communication Systems Icspcs 2024 Proceedings, 2024 Medical image segmentation plays a critical role in diagnosing and treating disease; it is a challenging task due to the complexity and variety of medical images. One of the critical factors in enhancing the accuracy and reliability of segmentation results is denoising, which involves removing noise from input images. Deep Neural Networks (DNNs) have shown promising results in medical image segmentation, but the manual design of DNN models is time-consuming and requires domain knowledge. To address this problem, Neural Architecture Search (NAS) provides an automated approach for developing neural network structures, eliminating manual design requirements. This paper introduces two evolutionary DNN models for medical image denoising and segmentation. The denoising model serves as a preprocessing step for segmentation, enabling us to assess its impact on medical image segmentation performance compared to normal images. The proposed approach performed well, with Dice scores of 86.99% and 88.62% on the original images for the Spleen and Heart segmentation datasets, respectively. While it achieved Dice scores of 89.12% and 87.78% on denoised images for the Heart and Spleen segmentation datasets, respectively. The experimental findings show that the architecture derived from the proposed approach outperforms existing models.
