Parameter-efficient deep-learning-based model for segmentation with radiomic feature extraction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Dynamic contrast-enhanced MRI (DCE-MRI) is crucial for breast cancer assessment.
• Accurate 3D tumor segmentation aids diagnosis, monitoring, and treatment planning.
• Current models like nnU-Net are computationally intensive, limiting their practical application.

Purpose of the Study:

• To propose a parameter-efficient convolutional neural network (CNN) architecture for breast tumor segmentation in DCE-MRI.
• To develop a model that reduces computational cost and memory requirements while maintaining high segmentation accuracy.
• To explore the prediction of treatment response to neoadjuvant chemotherapy.

Main Methods:

• Integration of lightweight residual blocks into a SegResNet backbone.
• Training on the first 3 DCE-MRI phases and testing the addition of FRLoss.
• Utilizing an encoder-decoder design for segmentation and exploring treatment response prediction with an XGBoost model.

Main Results:

• The proposed model achieved comparable performance to nnU-Net with a 0.99% higher Dice score.
• Significant reductions in parameter count (91.5%), FLOPs (85.05%), and memory usage (31.94%) compared to nnU-Net.
• The XGBoost model for treatment response prediction showed limited competitive performance (balanced accuracy 57.2%).

Conclusions:

• Parameter-efficient models can achieve competitive performance for DCE-MRI tumor segmentation.
• The developed model offers a more efficient alternative to nnU-Net for clinical applications.
• Further research is needed to improve treatment response prediction models.