Kidney Pathology Segmentation

Comparative study of CNNs and Transformer models for CKD gland segmentation using MONAI

Overview

A deep learning project comparing CNN and Transformer-based architectures for segmenting kidney glands in chronic kidney disease (CKD) pathology images.

Key Features

  • Model Comparison: Trained and evaluated multiple architectures including UNet, UNetR, and Swin UNetR
  • MONAI Framework: Leveraged MONAI (Medical Open Network for AI) for medical image analysis
  • Pathology Focus: Specialized in CKD gland segmentation from histopathology images
  • Architecture Analysis: Compared performance of CNNs vs. Transformer-based models

Models Evaluated

Model Architecture Type Key Features
UNet CNN Classic encoder-decoder with skip connections
UNetR Transformer Vision Transformer encoder with CNN decoder
Swin UNetR Transformer Shifted window attention for efficiency

Technologies Used

  • Framework: MONAI, PyTorch
  • Languages: Python
  • Tools: Slurm (HPC), Docker

Key Learnings

  • Transformer models showed promising results for medical image segmentation
  • MONAI provides excellent tools for medical AI development
  • Comparison insights useful for selecting appropriate architectures for pathology tasks