KETHANA09/TB-Detection-using-MobileNet

This project presents a lightweight, accurate, and portable deep learning model for detecting Tuberculosis (TB) from chest X-ray images using the MobileNet architecture and transfer learning. Designed to run efficiently on low-resource devices, this model can assist in fast, reliable, and accessible medical diagnosis. A Fine-Tuned MobileNet-Based Deep Learning Model for the Automated Diagnosis of Tuberculosis using Chest X-rays This repository contains the source code and models for a mini-project focused on developing a portable and effective deep learning model for the automated diagnosis of Tuberculosis (TB) using chest X-ray images. Project Overview Tuberculosis remains a significant global health issue. Early and accurate diagnosis is crucial, especially in resource-constrained environments. This project addresses this challenge by leveraging the MobileNet architecture with transfer learning to create a lightweight model suitable for deployment on devices with limited processing power. The model is fine-tuned on a dataset of TB-specific chest X-ray images from various open sources, including NIH, Montgomery County, Shenzhen Hospital, and Kaggle. Data augmentation techniques were employed to enhance image diversity and improve model performance. The repository includes: Source Code: Python code for data augmentation, preprocessing, model building, training, and evaluation for all the three models. Three different MobileNet-based models were explored: MobileNet + Dense Layer (Sigmoid Activation) MobileNet + 2 Dense Layers (ReLU) + 1 Dense Layer (Sigmoid) MobileNet + Convolution Layer with Max Pooling + 1 Dense Layer (Sigmoid) Dataset Details: kaggle.com Technologies Used: Python Tensorflow Pandas/NumPy Keras Matplotlib Seaborn Performance The project evaluated the performance of the different MobileNet models using metrics such as Accuracy, Precision, Recall, and F1-Score. The report provides detailed performance metrics for two different datasets. The model using pre-trained MobileNet weights with a single dense layer and sigmoid activation achieved high accuracy and outperformed more complex CNN models with fewer parameters, making it suitable for low-resource devices. Conclusion This project de …