Resume Projects

Handwriting Generation with style

Datasets:

• IAM Handwriting Database
• IAM On-Line Handwriting Database

WIP

Knee Osteoarthritis Detection

Dataset: Knee Osteoarthritis Dataset with Severity Grading

Project Overview

This project focuses on developing deep learning models to detect and classify knee osteoarthritis using medical imaging. By leveraging transfer learning and explainable AI techniques, I aim to create accurate and interpretable models for assisting in medical diagnoses. It demonstrates the application of advanced deep learning techniques in medical imaging, with a focus on balancing model performance with interpretability for real-world medical applications.

Key Features

• Classification Task: Detect and classify knee osteoarthritis on a scale of 0-3 (No/Minimal to Advanced)
• Deep Learning Models: Implemented using transfer learning with DenseNet, EfficientNet, EfficientNetV2, and ConvNeXt architectures
• Performance Metrics: Prioritized F2-score and Recall for medical relevance
• Dataset: 8.262 samples with class imbalance (60% in class 0)

Methodology

1. Data Preparation:

   • Simplified the original 0-5 scale to 0-3 for improved model performance
   • Addressed class imbalance using undersampling and data augmentation techniques

2. Model Development:

   • Utilized transfer learning with pre-trained models
   • Experimented with fine-tuning entire models vs. classifier layers only

3. Training and Optimization:

   • Implemented learning rate schedulers to mitigate overfitting
   • Explored various hyperparameters to optimize performance

4. Evaluation:

   • Primary metrics: F2-score and Recall
   • Secondary considerations: Confusion matrix and validation loss

Results

• Best model performance: 0.77 F2-score, 0.78 Recall
• Improved results observed when fine-tuning the entire model architecture

Future Work

• Further explore techniques to handle class imbalance
• Further experimentation with models.
• Check other models for this project like U-Net.

Wine Quality Prediction

Dataset: Wine Quality Dataset

Project Overview:

Experimented with 8 different machine learning models to predict wine quality on a 0-5 scale based on physicochemical properties. This project explores the application of various data science techniques to solve real-world problems in the food and beverage industry.

Key Challenges:

• Small dataset size, limiting the model's ability to generalize
• Highly imbalanced classes, potentially leading to biased predictions
• Complex relationship between chemical properties and perceived quality

Methodology:

1. Exploratory Data Analysis (EDA):

   • Conducted comprehensive statistical analysis to understand data distributions
   • Visualized correlations between features and wine quality
   • Identified potential outliers and their impact on the model

2. Data Preprocessing:

   • Implemented data normalization techniques to ensure consistent feature scaling
   • Applied oversampling methods (e.g., RandomOverSampler) to address class imbalance
   • Performed feature selection to identify the most influential wine characteristics

3. Model Development:

   • Experimented with various machine learning algorithms (e.g., Random Forest, XGBoost, SVM)
   • Utilized cross-validation
   • Implemented hyperparameter tuning through grid search

4. Performance Optimization:

   • Iteratively refined the model based on performance metrics (F1-score, Recall, Precision)
   • Analyzed feature importance to gain insights into key factors affecting wine quality