Self-Supervised Learning and Cross-Attention for Accurate Nasal Fracture DetectionIn Progress

Behrad Sadeghi, Ali Nabipour, Seyed Abolghasem Mirroshandel

Co-authored a research article in collaboration with Dr. Mirroshandel that employs cutting-edge models, including Vision Transformer (ViT) and EfficientNet, to diagnose nasal fractures using an original dataset that has not been explored in prior studies. We began by leveraging transfer learning, achieving over 85% accuracy and demonstrating that efficient models are effective for this medical imaging task. Building on this, I proposed the use of cross-attention to extract and integrate features from left and right nose views, given the paired nature of the images. To address the challenges associated with using pre-trained weights from non-medical datasets like ImageNet, we implemented self-supervised learning using SimCLR, significantly improving the model's performance. Our current work involves conducting comprehensive evaluations and applying and comparing various transformer architectures to rigorously document performance metrics.

PotholeSegmentation

This project focuses on the development and application of an advanced deep learning model, specifically YOLOv9, for the detection and segmentation of potholes in both images and video streams. Leveraging a custom dataset, the model has been fine-tuned to achieve high precision and recall rates through the use of optimized hyperparameters, data augmentation techniques, and advanced training strategies such as mixed precision training and gradient accumulation. Key achievements include a mAP50 of 0.807 and a Mask mAP50 of 0.825, underscoring its effectiveness in practical applications.

Amazon Reviews

A machine learning model was created and trained on Amazon Reviews to perform sentiment analysis. The dataset underwent comprehensive preprocessing steps, including tokenization, removal of stopwords, lemmatization, stemming, elimination of tags and emojis, and normalization. Logistic Regression was utilized for the sentiment analysis task on Amazon Reviews. The repository includes resources and code for implementing various sentiment classification models, such as Embedding Models, BERT, and Simple Neural Networks.

AlzMRI-Net

The AlzMRI-Net project develops a deep learning model to classify MRI scans into four stages of Alzheimer's disease: MildDemented, ModerateDemented, NonDemented, and VeryMildDemented. Leveraging transfer learning with a pre-trained EfficientNet-V2-L model, this work fine-tunes the model on a specialized dataset using PyTorch. Advanced techniques such as mixed precision training and gradient accumulation are employed to enhance performance and efficiency. The model achieves a test accuracy of 99.19%, with comprehensive evaluation metrics, including precision, recall, and AUC, underscoring its effectiveness in Alzheimer's disease classification.

Paper Summarizer

The goal of this project is to simplify the process of reviewing the latest research papers by providing concise summaries generated using BART, a state-of-the-art large language model (LLM). The tool fetches the most recent papers from PubMed and arXiv based on a user-defined query and presents them through an interactive interface. Users can easily select papers and view detailed summaries, making it easier to stay updated with current research.

Credit Card User Clustering

This project involves clustering and segmenting around 9,000 active credit card users based on their behavior over a six-month period, using 18 behavioral variables. Customer data is preprocessed and reduced to two dimensions via Principal Component Analysis (PCA). K-Means clustering is then applied to group customers into four distinct clusters, enabling analysis of their behaviors and identification of cluster-specific characteristics.

Predicting-Loan-Acceptance

In this project, I had access to a bank dataset to predict which individuals the bank should target with personal loan offers based on historical data. I also employed machine learning models like Logistic Regression, Naive Bayes, and K-Nearest Neighbors, and assessed loan acceptance probabilities.

Smart Railway System

This project is my final assignment for the Data Structures course. It involves developing a Smart Railway Transportation System using graph data structures and algorithms. The system addresses challenges such as finding the shortest path between stations, detecting cycles in railway routes, identifying all possible paths between stations, and determining maximal cliques within the network. We utilize algorithms like Dijkstra’s for shortest paths and DFS for pathfinding and cycle detection.

Course-Registration-System

For my Advanced Programming course, I developed an Integrated University Course Registration System using object-oriented principles. The project included a detailed UML diagram to represent the system’s architecture and interactions.

Four In Row

The game is designed for single-player mode, featuring an AI Agent that employs the Negamax algorithm. This algorithm evaluates the minimum values of child nodes in each step and then chooses the maximum of these minimum values as the value for the parent node in the subsequent step.