Featured Projects

🤝 Social Decision-Making and Helping Behavior: Can Q-Learning Models Learn to Help? 

• Developed computational models using Q-Learning and Deep Q-Learning to simulate helping behaviors in a two-player game, inspired by the Farm Task from Osborn Popp and Gureckis (2024).

• Analyzed how reciprocity, costs, resource disparities, and visibility influence decision-making, incorporating temporal difference learning and neural networks to model dynamic interactions.

• Findings revealed reciprocity as a key driver of helping behaviors, with environmental factors like energy costs also playing significant roles, advancing insights into altruistic behavior and cooperative AI systems.

🧪 Region of Interest Based Medical Image Compression

•  Explored Region of Interest (ROI) coding for medical image compression, leveraging UNET segmentation to identify tumor regions in the BRaTS 2020 dataset and applying HEVC compression to optimize storage and transmission.

• Achieved efficient compression with preserved diagnostic quality by prioritizing tumor regions while compressing less critical areas more aggressively, meeting telemedicine demands.

• Research contributed to balancing compression rate and image quality, paving the way for advanced solutions in large-scale medical imaging and remote healthcare services.

🩺  U-Trans: An Automated Swin-Transformer based model for Skin Lesion Segmentation

• Developed a U-SwinTransformer model integrating DCT, DWT, and FFT preprocessing, achieving 95.91% accuracy, 88.41% Dice Coefficient, and a 0.1042 final loss on the ISIC 2017 dataset for skin lesion segmentation.

• Enhanced global feature extraction in medical image segmentation by embedding Swin Transformer Blocks into a U-shaped architecture, improving the Jaccard Index to 79.46% and reducing validation loss to 0.1370.

• Research Submitted for publication.

🚀 RxVision: Reducing Medical overdose and underdose using visual tracking

• Launched RxVision at Hacklytics 2024 (Georgia Tech), tackling drug underuse and overuse with a focus on preventing opioid overdose deaths, which claim over 130 lives daily in the U.S.

• Developed recognition technology to validate drug-patient matches and monitor dosage schedules, minimizing medication errors and improving patient safety.

⚡Transfer Learning Approach to optimize 3D Brain MRI Segmentation

• Developed advanced 3D segmentation for brain tumor detection using a U-Net architecture and HPC techniques, reducing training time per slice by 317%.

• Optimized computational efficiency through data parallelism and transfer learning, achieving a 22% improvement in segmentation accuracy for precise and scalable medical image analysis.

🔬SkinCheck: An AI-Powered Skin Cancer Classification Tool

• Developed a Classification Tool using 40,000 augmented images, achieving 87.89% accuracy through deep learning experimentation to identify the most effective algorithm. Designed an intuitive graphical interface to enable easy detection and classification of skin cancer types, ensuring accessibility for all users.

• Authored and Published a paper titled “A Transfer Learning-based GUI for Skin Cancer Diagnosis and Classification using Dermoscopic Images” to the IEEE SILCON 2023 conference in India.