MED-CXX

MED-CXX is a C++ project built with an object-oriented design, aiming to solve key medical imaging problems through deep learning. The project focuses on implementing and training several convolutional neural network (CNN) architectures — including U-Net, ResNet, and DenseNet — entirely in C++ using LibTorch (the PyTorch C++ API).

These models are applied to core medical tasks such as blood vessel segmentation, disease detection from medical scans, and image-based classification of pathological conditions.

MED-CXX was developed as part of the Object-Oriented Programming (OOP) lab course project at the Faculty of Mathematics and Computer Science, University of Bucharest, with a focus on clarity, reusability, and clean class structure.

---

Changelog

Implemented Features (May 2025) (v0.2)

• UNet, ResNet, and DenseNet implemented using LibTorch
• Blood vessel segmentation on DRIVE/STARE datasets with visualization
• Weighted BCE + Dice loss for handling class imbalance
• Model saving/loading utilities (saveModel, loadModel)
• GIF/video demo generator for showing side-by-side input / ground truth / prediction
• ImageLoader with caching, preprocessing, and grayscale/threshold conversion
• Modular class structure (BaseModel, ImageLoader, Benchmark, etc.)
• Metrics: Accuracy, Precision, Recall, F1 Score, IoU, MAE, Hausdorff
• Auto device selection (CUDA or CPU)

Implemented Features (June 2025) (v0.3)

• Unified med-cxx runner CLI via common::ArgParser
• Config flags for:
  • --model : UNet / DenseNet / ResNet
  • --train-dir : path to training images (and masks for segmentation)
  • --test-dir : path to test images (and masks)
  • --cls-train-dir: list of class-folders for classification
  • --model-name : human-friendly name (prefix for weight files & demos)
  • --weights : load existing .pt model file
  • --skip-train : skip training if weights are provided
  • --epochs, -e : number of training epochs
  • --lr : learning rate
  • --bce-weight : positive‐class weight for BCE loss
  • --cuda : request GPU (falls back to CPU if unavailable)
  • --resnet-ver : select ResNet18/34/50/101/152
  • --video : generate demo video (segmentation)
  • --fps : frames per second for demo video
  • --hold : how many frames to hold each sample
  • --bar-width : width of the training progress bar
• Extracted BaseTrainer, SegmentationTrainer, ClassificationTrainer
• printProgressBar utility in common::Utils
• Clean separation: runners/, trainer/, common/, models/
• Design Patterns:
  • Template Method
    BaseModel::predict defines the skeleton of inference, sub-classes (UNet, ResNet, DenseNet) fill in the details.
  • Composite
    All of our layer modules (DoubleConv, Down, Up, BasicBlock, etc.) inherit from a common base and can be nested arbitrarily inside Sequential containers.
  • Factory
    ArgParser + runner’s switch(cfg.modelType) cleanly instantiates different BaseModel subtypes.
  • Singleton (planned)
    A single global Config object to hold all CLI flags and share across trainers/visualizers.
  • Facade (planned)
    BaseTrainer + SegmentationTrainer/ClassificationTrainer hide all the details of data loading, optimization, loss, metrics and video generation behind a simple train() / evaluate() interface.

---

Usage examples

# Show help
./med-cxx --help

# Segmentation: train & evaluate UNet 
./med-cxx \
  --model UNet \
  --train-dir data/train \
  --test-dir  data/test \
  --model-name vessels \
  --epochs 200 \
  --lr 1e-4 \
  --bce-weight 3.0 \
  --cuda \
  --video \
  --fps 2 \
  --hold 3

# Classification: train & evaluate DenseNet on folder-per-class
./med-cxx \
  --model DenseNet \
  --cls-train-dir data/cls/train \
  --cls-test-dir  data/cls/test \
  --model-name  classification_run \
  --epochs 50 \
  --lr 5e-4 \
  --skip-train  \
  --weights  classification_run.pt

# Run ResNet50 inference only on CPU
./med-cxx \
  --model ResNet \
  --resnet-ver 50 \
  --cls-test-dir data/cls/test \
  --skip-train \
  --weights resnet50_run.pt

---

Installation

Linux

# Install dependencies
sudo apt update
sudo apt install cmake libopencv-dev build-essential

# Clone the repo
git clone https://github.com/IAMSebyi/med-cxx.git
cd med-cxx

# Download LibTorch (instead of <COMPUTE PLATFORM>, use cpu, cu118, cu126, cu128)
wget https://download.pytorch.org/libtorch/<COMPUTE PLATFORM>/libtorch-cxx11-abi-shared-with-deps-2.7.0%2B<COMPUTE PLATFORM>.zip
unzip libtorch-cxx11-abi-shared-with-deps-2.7.0%2B<COMPUTE PLATFORM>.zip

# Build the project
mkdir build && cd build
cmake ..
make

Windows

1. Install dependencies

• CMake
• OpenCV
• Visual Studio 2022

2. Clone the project

git clone https://github.com/IAMSebyi/med-cxx.git
cd med-cxx

3. Download LibTorch (CPU or CUDA 11.8, 12.6, 12.8)

• LibTorch

4. Configure and build using CMake

mkdir build && cd build
cmake ..
make

5. Ensure *.dll files from libtorch/lib and opencv/build/x64/vcXX/bin are either in your PATH or next to your executable.

Mac

# Install dependencies
brew install cmake opencv

# Clone the repo
git clone https://github.com/IAMSebyi/med-cxx.git
cd med-cxx

# Download LibTorch (only CPU supported)
https://download.pytorch.org/libtorch/cpu/libtorch-macos-arm64-2.7.0.zip
unzip libtorch-macos-arm64-2.7.0.zip

mkdir build && cd build
cmake ..
make

---

Resources

• PyTorch Documentation
• OpenCV Documentation
• U-Net: Convolutional Networks for Biomedical Image Segmentation - Olaf Ronneberger, Philipp Fischer, Thomas Brox
• Deep Residual Learning for Image Recognition - Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
• Densely Connected Convolutional Networks - Gao Huang, Zhuang Liu, Laurens van der Maaten, Kilian Q. Weinberger
• Locating blood vessels in retinal images by piecewise threshold probing of a matched filter response - Adam Hoover, V. Kouznetsova, Michael Goldbaum