Recent Pre-prints & Publications

• 

  Performance Analysis of LLMs for Medical Text Summarization

  Abstract

  Due to the rapid expansion of medical literature, keeping pace with the latest research and clinical guidelines has become more challenging for healthcare professionals. To overcome this challenge, effective text summarization is crucial for improving access to knowledge, enhancing clinical decision-making, and ultimately benefiting patient outcomes. In this study, a medical text summarization system that employs large language models (LLMs) was fine-tuned and evaluated with the objective of generating precise, logical, and brief summaries of medical literature, emphasizing clinical relevance and ease of understanding. We plan to evaluate the performance of GPT3, GPT4 and the fine-tuned T5, BART, and Pegasus models trained on the standard PubMed dataset using standard evaluation metrics.
  
  Link

• 

  Medical Image Enhancement using GANs Architecture

  Abstract

  Medical Imaging is used by radiologists for diagnostic purposes and to check for abnormalities, and these imaging techniques involve radiation. Overexposure to radiation can have an adverse impact on the human body, and using less radiation gives us a noisy output. Hence, radiologists find it difficult as there is a trade-off between the amount of radiation that can be used and the quality of the image. Moreover, noise in medical images can occur due to fluctuation of photons, a reflection of radiations from the subject, or due to instrumental vibration or faults. The proposed approach is a pipeline which starts with denoising using GANs architecture, in which two models have been trained, one for handling all kinds of noise and the second one specifically for Poisson noise. Further, post-processing methods like single-shot HDR using Retinex Filtering and Edge Enhancement using unsharp masking have been done to get a structurally more similar and enhanced denoised image.
  
  Link

• 

  Empirical Study of Supervised and Active Learning Methodologies

  Abstract

  Document classification is one of the predominant tasks in Machine Learning, that by and large involves the use of supervised methodologies to determine most suitable categories for a given set of documents. In this project, we seek to conduct an empirical study, that, in addition to the conventional machine learning methods, experiments with active learning and semi-supervised methods for assorting a mix of labelled and unlabelled computer science scholarly articles hosted on the arXiV repository. We develop a solution using the gathered labelled data points, by experimenting with various multi-class supervised learning models such as K Nearest Neighbours, Naive Bayes, Random Forest, and Logistic Regression. We also account for the issue of class imbalance, and thus conduct experiments to determine a suitable oversampling methodology for addressing this concern. Lastly, we extend beyond the conventional machine learning classification methods to incorporate the concept of active learning, a type of iterative supervised learning, used when the unlabelled data is in abundance.
  
  Link