Detection of Brain Mid-Sagittal Plane Based on Progressive Semi-Supervised Pixel Classification Algorithm.

Abstract

Automatic detection of mid-sagittal plane (MSP) in the brain is widely used for symmetry analysis, midline shift (MLS) measurement, tilt correction, and brain morphometry. Existing MSP detection methods typically employ fully supervised learning (SL). However, this approach is greatly constrained by the quantity and quality of expertly annotated data. Due to the high cost of annotation, existing methods can only use a limited number of samples with a single type for model training, resulting in low generalization of such models. We propose an MSP detection framework to improve the model's generalization performance across various types of data. Specifically, on one hand, we design a Progressive Semi-supervised Learning (PSSL) method based on the morphological characteristics of MSP, enabling the model to achieve continuous performance improvement from a large amount of unlabeled data. On the other hand, we incorporate a correction mechanism into the model using neighborhood information in three-dimensional space, providing the model with a certain degree of fault tolerance. Extensive validation conducted on five datasets (comprising millions of brain sections) indicates that our method outperforms state-of-the-art approaches in midline detection within the brain.