medmodeler

This is the code for medmodeler, an end to end automatic segmentation tool for MRI/CAT files. This model identifies various tissues like lung, bladder, liver, kidney and bone, and then produces a medical image file for ease of visualization using torchio.

Usage

For use in your own project, simply download the project, install its dependencies and import "inference()" from inference.py file

Examples

Architecture

• This project utilizes the famous PyTorch segmentation repo by qubvel that can be found here
• Specifically, this model uses the Feature Pyramid Network Architecture, with the Mix Vision Transformer as the encoder and pre-training performed on ImageNet

Dataset

• This model was trained on the CT-ORG Dataset

Room for Improvement

• The model performs well on bone and liver segmentation, but often fails to fully capture the other classes, most likely due to insufficient training
• When inferencing, the tensor_to_label_map() function is a huge bottleneck, and could likely be implemented more efficiently with matrix operations
• No input validation, inference only accepts traditional medical image formats like *.nii.gz

Credit

This project was produced in collaboration with the University of Florida, who supplied compute resources in the form of 4 NVidia A100 GPUs.