ORCHESTRAL FULLY CONVOLUTIONAL NETWORKS FOR SMALL LESION SEGMENTATION IN BRAIN MRI

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Neurology

Background:

• White matter (WM) lesion segmentation is crucial for diagnosing neurological conditions and predicting outcomes.
• Existing Convolutional Neural Networks (CNNs) struggle with segmenting small, deep WM, and sub-cortical lesions.
• Anemia can be associated with white matter abnormalities requiring accurate segmentation.

Purpose of the Study:

• To develop and evaluate a novel deep learning approach for accurate segmentation of small white matter lesions.
• To address the limitations of current methods in detecting subtle WM lesions.
• To improve the management of unbalanced data in lesion segmentation tasks.

Main Methods:

• Utilized multi-scale and supervised fully convolutional networks (FCNs).
• Applied the proposed method to segment small WM lesions in a cohort of 22 anemic patients.
• Implemented a multi-supervised scheme to handle data imbalance.

Main Results:

• The proposed multi-scale, supervised FCN achieved a Dice score of 0.78 on the testing dataset.
• This represents a significant improvement over a single FCN, which yielded a Dice score of approximately 0.31.
• The multi-scale approach effectively identified small lesions and reduced false positives.

Conclusions:

• Multi-scale and supervised FCNs offer a promising solution for accurate segmentation of small white matter lesions.
• This technique enhances diagnostic capabilities for neurological conditions associated with WM abnormalities.
• The developed method shows potential for clinical application in managing patients with anemia and related neurological findings.