Optimizing the use of radiologist seed points for improved multiple sclerosis lesion segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• Multiple sclerosis (MS) lesion segmentation often relies on radiologist input, which can be inefficient and assumes optimal placement.
• Current methods may not offer an intuitive workflow for experts placing seed points.

Purpose of the Study:

• To develop and evaluate intuitive heuristics for processing seed points in MS lesion segmentation.
• To improve segmentation accuracy and facilitate natural seed point placement for enhanced efficiency.

Main Methods:

• A Parzen window-based classifier was used to automatically delineate MS lesions based on radiologist-placed seed points.
• Two radiologists provided seed points on a large set of MS clinical trial MRIs.
• An interactive region-growing method generated ground truth for evaluation.
• Dice coefficient (DC) and Spearman’s rank correlation were used to measure segmentation agreement.

Main Results:

• Heuristics significantly improved DC (+32.91 pts) and correlation (+0.50) for low lesion load scans.
• Moderate improvements were observed for medium lesion load scans (DC +14.55 pts, correlation +0.15).
• High lesion load scans showed minimal improvement due to already accurate segmentation by Parzen windows.
• With heuristics, DC approached 80% and correlation exceeded 0.9 across all lesion load categories.

Conclusions:

• The proposed intuitive heuristics enhance the accuracy and efficiency of MS lesion segmentation.
• These methods are particularly beneficial for scans with lower lesion burdens.
• The approach offers a more natural and effective workflow for radiologists performing lesion segmentation.