Efficient, high-performance semantic segmentation using multi-scale feature extraction.

Area of Science:

• Medical Imaging
• Deep Learning
• Computer Vision

Background:

• Deep learning success relies on large datasets, which are limited in medical applications due to data sensitivity.
• Federated learning enables collaborative machine learning but requires efficient models for network data transfer.

Purpose of the Study:

• To develop a small, highly optimized neural network for medical image segmentation suitable for resource-constrained environments.
• To evaluate MoNet's performance in clinical tasks and its suitability for privacy-preserving collaborative learning.

Main Methods:

• Introduced MoNet, a shallow, U-Net-like architecture using dilated convolutions with decreasing rates.
• Applied and tested MoNet on pancreatic CT and brain MRI tumor segmentation tasks.
• Assessed segmentation performance, generalization, parameter efficiency, and inference latency on CPU.

Main Results:

• MoNet achieved performance on par with larger architectures on segmentation tasks.
• Demonstrated superior out-of-sample generalization and significantly fewer parameters.
• Showed substantial reduction in model storage, confirming suitability for federated learning.
• Confirmed utility in environments without GPUs through CPU inference latency evaluation.

Conclusions:

• MoNet is an efficient deep learning architecture for medical image segmentation, suitable for resource-constrained and federated learning settings.
• The model offers competitive performance with improved generalization and reduced computational/storage requirements.
• Publicly available open-source implementation facilitates broader adoption.