Enhancing Segmenter with Masked Patch Modeling: Development of the Mask-Segmenter Model
Vaughn Nephi F. Fajardo, Karl Raphael L. Paradeza, Raphael Alampay, Patricia Angela R. Abu · 2024
Transformer models have shown great potential in segmentation of MRI scans due to the ability to use contextual information to establish relationships. However, these models have been susceptible to noise. This study introduces Mask-Segmenter, an enhancement to the Segmenter model, aiming to improve MRI scan segmentation and noise resilience by utilizing masked patching and Sobel Edge Detection. This technique highlights significant image edges to identify key patches for masking, allowing the model to predict masked areas using contextual information from adjacent patches, using a learnable threshold. Sobel Edge Detection was applied during training to identify significant image edges by analyzing pixel-intensity gradients, guiding the selection of patches for masking based on a set threshold. The model then predicted the labels of these masked patches using contextual information from adjacent unmasked areas, producing contextualized encodings that a decoder used to create class masks and 2D feature maps. The model was evaluated using the BraTS dataset, augmented with Gaussian noise, comparing its performance to State-of-the-art models. Mask-Segmenter demonstrates significant improvement over Segmenter, SETR, and ResNet-101 in brain MRI segmentation, while Swin L-UperNet's performance identifies areas for Mask-Segmenter's further enhancement in handling noise and detail. The findings affirm Mask-Segmenter's potential in clinical applications.