Combining simple interactivity and machine learning: a separable deep learning approach to subthalamic nucleus...

Area of Science:

• Neurosurgery
• Medical Imaging
• Machine Learning

Background:

• Deep brain stimulation (DBS) is a crucial treatment for neurological disorders like Parkinson's disease.
• Accurate preoperative planning, including precise segmentation of targets like the subthalamic nucleus (STN), is vital for effective DBS electrode placement.
• Segmenting small, low-contrast structures such as the STN in MRI is challenging, exacerbated by issues like class imbalance in machine learning.

Purpose of the Study:

• To develop and evaluate a novel two-stage separable learning workflow for automated STN segmentation.
• To improve the accuracy and efficiency of STN segmentation compared to current clinical methods.
• To enhance the clinical utility of STN segmentation through an interactive correction mechanism.

Main Methods:

• A two-stage workflow was implemented: first, a localization step to detect and crop the STN region, followed by a segmentation step within the cropped area.
• The performance of the proposed method was compared against a standard registration-based technique.
• A human-computer interaction experiment was conducted with novice and expert users to assess the effectiveness of the interactive correction feature.

Main Results:

• The two-stage separable learning method significantly outperformed the current registration-based approach for STN segmentation.
• The segmentation accuracy approached the limits imposed by image resolution.
• The interactive correction mechanism, enabled by the two-stage design, substantially improved user performance and error correction capabilities.

Conclusions:

• Separable learning offers a feasible and effective approach for fully automatic STN segmentation.
• The proposed method enhances interactivity, facilitating its adoption in clinical practice for DBS planning.
• This workflow represents a significant advancement in automating and improving the precision of neurosurgical planning.