Enhancing nnUNetv2 Training with Autoencoder Architecture for Improved Medical Image Segmentation.

Area of Science:

• Radiotherapy and Oncology
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Accurate auto-segmentation of gross tumor volumes (GTVs) in head and neck cancer (HNC) is crucial for effective MRI-guided radiotherapy (RT).
• Current segmentation methods face challenges in precision, impacting clinical workflows.

Purpose of the Study:

• To develop and evaluate a novel deep learning model for enhanced auto-segmentation of GTVs in HNC using MRI-guided RT images.
• To improve the accuracy and efficiency of tumor delineation in radiation oncology.

Main Methods:

• Development of a modified nnUNetv2 deep learning framework incorporating an autoencoder architecture.
• Inclusion of original training images as an additional input channel and utilization of Mean Squared Error (MSE) loss function.
• Training on 150 HNC patient datasets and private evaluation on 50 test patients for the HNTS-MRG 2024 challenge.

Main Results:

• Achieved an aggregated Dice Similarity Coefficient (DSCagg) of 0.8516 for metastatic lymph nodes (GTVn).
• Obtained a DSCagg of 0.7318 for the primary tumor (GTVp), with an average DSCagg of 0.7917 across both structures.
• Demonstrated that the enhanced nnUNet architecture effectively learned additional image features, improving segmentation accuracy.

Conclusions:

• The modified nnUNetv2 framework with an autoencoder and combined loss functions significantly enhances auto-segmentation accuracy for HNC in MRI-guided RT.
• This deep learning approach contributes to more precise and efficient clinical workflows in radiation oncology.
• The model shows promise for improving treatment planning and delivery in head and neck cancer radiotherapy.