Ahmed Abdullah
Experience
Computational Oncology Research: Developing a predictive model for tumor de-growth using the Gompertz function to assess drug efficacy. Implementing mathematical modeling and data-driven approaches to analyze tumor response, aiming to improve accuracy in treatment outcome predictions. Working with oncological datasets and statistical techniques to enhance precision medicine strategies.
Teaching Assistant for Computer Organization & Assembly Language (EE2003) and Object-Oriented Programming (CL1004). Conducted lab sessions, guided students through complex programming concepts, and assisted in debugging and problem-solving. Evaluated assignments, held office hours, and collaborated with faculty to improve student learning experiences.
Multimodals Research: Worked on handling missing modality and accurately estimating face-voice association using FOP Loss. Proposed a novel approach leveraging WavLM, ECAPA, and VGGVox on the VoxCelebV1 dataset to improve embedding quality through pseudo-modal and zero-shot learning. Addressed challenges in cross-modal learning and optimized the model for robustness in real-world scenarios. Achieved an Equal Error Rate (EER) of 20%, outperforming previous benchmarks in voice-face matching accuracy.
Healthcare Research: Worked on a novel approach using Pathopix-GANs for CT scan augmentation in ischemic stroke functional outcome prediction. Addressed data scarcity (only 43 patients available) by enhancing multimodal imaging. Utilized the ISLES’24 dataset and surpassed the previous state-of-the-art accuracy of 43%.
Generative AI / NLP Research: Worked on Large Language Models (LLMs) and their impact on gender and race biases in image generation. The project involved analyzing the biases in AI-generated imagery and developing methods to mitigate these biases. My work helped improve the ethical use of AI in generating fairer, more diverse imagery in applications like advertising, media, and content generation.
Software Development (MERN Stack): Contributed to the development of web applications using the MERN stack (MongoDB, Express, React, Node.js). I worked on building responsive front-end features and improving back-end functionality. This experience helped me strengthen my full-stack development skills and provided me with hands-on exposure to the software development lifecycle, debugging, and optimization techniques.
Manuscripts
This project addresses the classification of degenerative lumbar spine conditions using GANs to impute missing MRI data. Over 70 models were tested, achieving state-of-the-art results with a minority class accuracy of 92.24% and an AUC-ROC of 64%. The approach effectively handles incomplete MRI datasets, leveraging GANs for data imputation and robust classification. The study demonstrates significant improvements in classifying underrepresented conditions. It highlights the potential of advanced generative and classification techniques in medical imaging.
This study examines the rise in research paper retractions and declining research quality in Pakistan through surveys and interviews with over 300 participants. Key issues include plagiarism, data fabrication, lack of ethics training, institutional barriers, and the pressure to “publish or perish.” Results reveal that international aid supports less than 60% of researchers, and 70% of students lack awareness of research ethics. To address these issues, the study develops a cosine similarity tool for citation analysis, reducing errors in academic writing. It recommends mandatory ethics training, institutional support, and automated tools to promote ethical research and improve Pakistan’s academic reputation.
Projects
A state-of-the-art model to classify degenerative MRIs for the Lumbar Spine with an ability to impute the missing MRI using the pseudo-modality approach.
AskFAST is a chatbot designed to handle admission-related queries for FAST, streamlining the process with AI-driven responses.
Developed a multimodal system to detect skin cancer, combining cutting-edge CNNs with detailed image analysis to elevate skin cancer detection.
The LBW (Leg Before Wicket) Review System detects and analyzes ball movements in cricket matches, specifically focusing on LBW scenarios.
Developed a multimodal search engine using CLIP by OpenAI with Flask API for backend and HTML/CSS for the frontend web application, using image embeddings for searching images.
Created an interactive candlestick chart in C++ to visualize financial market data with an emphasis on stock price movements.
Developed a model to predict survival outcomes for Hematopoietic Cell Transplantation (HCT) patients, leveraging patient data and machine learning techniques.
AutoCiter is a Flask-based tool that uses sentence transformers and cosine similarity to automatically cite text from uploaded PDF papers.
This project recommends recipes to users based on their preferences using PySpark and GPU acceleration, leveraging distributed computing capabilities.
This project performs sentiment analysis on Twitter data in a batch processing manner using the ntscraper library, Hadoop clusters (HDFS), and PySpark.