This project aims to compare and evaluate the performance of three popular models for image segmentation—U-Net, FCN (Fully Convolutional Network), and DeepLabv3—specifically in the task of lung image segmentation by instance. The goal is to accurately segment lung regions in medical images, distinguishing individual instances of lungs.
The U-Net model, known for its U-shaped architecture, has demonstrated success in various segmentation tasks. FCN, on the other hand, utilizes fully convolutional layers to achieve end-to-end segmentation. DeepLabv3 is a state-of-the-art model that incorporates atrous convolution and dilated convolutions to capture fine details in images.
The project involves training and testing these models using a dataset of lung images labeled with precise instance-level segmentation masks. Performance evaluation will be conducted using metrics such as Intersection over Union (IoU), Dice coefficient, and accuracy. Additionally, qualitative analysis will be performed to assess the visual quality of the segmentation outputs.
The findings of this project will provide insights into the strengths and weaknesses of each model for lung image segmentation by instance. Such information can aid in selecting the most suitable model for similar medical imaging tasks, ultimately contributing to improved analysis and diagnosis in the field of healthcare.