This repository contains python codes related HEROVision model, which shows promise in optimizing individualized treatment between TA and SR for early-stage rHCC, complementing current clinical guidelines.The training of the model was completed on NVIDIA GeForce GTX 3090.
- The data required to build the model dataset.py
- Model architecture Model_hovertrans.py
- Training and evaluating surgical models train_hover_SR.py
- Training and evaluating ablation models train_hover_TA.py
- Some configuration information config.py
To run the codes ensure you have the necessary prerequisites:
- Python installed on your system (version 3.7 or above).
- Required libraries: torch, torchvision, numpy, SimpleITK, pydicom, lifelines, radiomics, matplotlib, and scipy. You can install them using pip:
63D0
pip install torch torchvision numpy SimpleITK pydicom lifelines radiomics matplotlib and scipy
dataset.py
- Description: The dataset comprises four modalities of medical images:
- Grayscale Ultrasound (US)
- Contrast-Enhanced Ultrasound (CEUS)
- T2-Weighted Imaging (T2WI)
- Diffusion-Weighted Imaging (DWI)
- Number of Images: 15774 images.
- Label: PFS.
- Splitting: The datasets were divided into a training cohort and an internal validation cohort in a 4:1 ratio, based on chronological order.
- Resizing: All images are resized to a uniform dimension (e.g., 128*128 pixels) to standardize input for the neural network.
- Normalization: Image pixel values are normalized to the range [0, 1] to facilitate faster convergence during training.
Model_hovertrans.py
Architecture: This architecture comprises four main components:
- The embedding layer, responsible for partitioning the original image into patches using diverse techniques;
- The feature extraction layer, leveraging the ViT model to extract and integrate features within each patch;
- Convolutional layers, which further integrate features extracted from patches using various partitioning methods;
- The Cox regression layer, utilizing the COX loss function to regress the extracted features and ultimately derive the corresponding risk value.
Hyperparameters: During the training phase, we resized the images to 128 * 128 dimensions. We employed two stages with depths of 4 and 2, heads of 2 and 4, and dimensions of 4 and 8, respectively. The Cox loss function was selected. We utilized the Adam optimizer with a learning rate of 0.001 and a decay rate of 0.0001. Training was conducted over 100 epochs with a batch size of 16.
train_hover_SR.py OR train_hover_TA.py
- Description:
- Training Procedure
- Optimizer: The Adam optimizer is used for its efficiency in training deep learning models.
- Loss Function: Cox Loss is used to regression time information.
- Callbacks: Early stopping and model checkpointing are employed to prevent overfitting and save the best model during training.
- You can train the model use instruction:
python train_hover_SR.pyorpython train_hover_TA.py
- We plan to apply the HEROVision model in a prospective study to optimize personalized curative treatment for early-stage recurrent hepatocellular carcinoma.