Logistic Regression-Based Model Is More Efficient Than U-Net Model for Reliable Whole Brain Magnetic Resonance...

Area of Science:

• Medical Imaging Analysis
• Computational Neuroscience
• Machine Learning in Medicine

Background:

• Automated whole brain segmentation is crucial for identifying volumetric markers in neurological diseases.
• Deep learning methods like U-Net show promise but struggle with limited training data.
• Manual segmentation is time-consuming, necessitating efficient training strategies.

Purpose of the Study:

• To compare the performance of U-Net and Classification using Derivative-based Features (C-DEF) for whole brain segmentation.
• To evaluate performance under conditions of limited training data availability.

Main Methods:

• U-Net and C-DEF models were trained on data from 5, 10, and 15 participants across HIV and multiple sclerosis cohorts.
• Performance was also assessed on a large brain tumor segmentation dataset.
• Dice similarity coefficient was used for statistical comparison.

Main Results:

• C-DEF achieved superior segmentation for lesion and cerebrospinal fluid classes with ≤15 training participants.
• U-Net demonstrated significant improvement with increased training data (24%-30%).
• C-DEF matched or exceeded U-Net for enhancing tumor and edema segmentation; U-Net excelled in necrotic tumor segmentation with ≥10 participants.

Conclusions:

• Classical machine learning (C-DEF) provides more accurate brain segmentation than deep learning (U-Net) with limited training data (≤15 participants).
• U-Net may be preferable for deep gray matter and necrotic tumor segmentation with larger datasets (≥20 participants).
• Classical methods like C-DEF can avoid manual annotation bottlenecks when retraining models for new imaging protocols or pathologies.