Towards Digital Twin of RF Ablation: Real-Time Prediction of Time-Dependent Thermal Effects Using Transformer

Overview

This repository contains the implementation of the neural network model presented in the MICCAI 2025 workshop [Digital Twin for Healthcare], as described in the paper "Towards Digital Twin of RF Ablation: Real-Time Prediction of Time-Dependent Thermal Effects Using Transformer."

Features

• A neural network model for prediction of RF ablation outcomes(damage area and temperature distribution) based on [UNETR: Transformers for 3D Medical Image Segmentation] by Ali Hatamizadeh, Dong Yang, Holger Roth, and Daguang Xu (2021).
• The damage prediction model uses Dice loss, while the temperature prediction model uses MSE loss. Both models share the same UNETR architecture, but their parameters are trained separately.

Requirements

• python 3.10.12
• cuda 11.8
• torch 2.0.1
• numpy 1.26.4

Dataset

• The composition of the dataset
  • input : a 3D tumor image and a 3D needle image
  • output : an 18-channel 3D image(damage area or temperature distribution) over time.
• The tumor image was constructed using the publicly available dataset Saha et al., 2018.

References

• 'UNETR.py' was designed with reference to UNETR implementation by tamasino52, which is licensed under the MIT License.

Contact

For any queries, please reach out to Seonaeng Cho. (seonaeng@yonsei.ac.kr)