MDD-Net: Medical Segmentation with Parallel Denoising Transformation
Sophia Brown, James Davis, Jannat Roy · Preprints.org · 2024
The automated segmentation of brain gliomas from multimodal MRI scans is pivotal in both clinical trials and everyday medical practice. Manual segmentation, however, presents significant challenges due to its labor-intensive nature, high costs, and susceptibility to inaccuracies. These difficulties arise despite the involvement of skilled human experts, primarily because of the substantial variability and uncertainty inherent in human annotations. Addressing these issues, our study introduces MDD-Net, an advanced end-to-end deep learning-based segmentation framework. This framework leverages a novel multi-decoder architecture that concurrently addresses three distinct sub-tasks through a partially shared encoder, enhancing the model's ability to generalize across different segmentation challenges. Additionally, we incorporate sophisticated smoothing techniques applied to the input MRI images, producing denoised versions that serve as supplementary inputs to the network. This dual-input strategy not only mitigates noise-related distortions but also enriches the feature representation, leading to more precise segmentation outcomes. Comprehensive validation results demonstrate a significant improvement in segmentation performance when employing the proposed MDD-Net method compared to existing approaches. This advancement holds promise for enhancing diagnostic accuracy and streamlining workflows in clinical settings. Future work will explore the integration of additional imaging modalities and the application of MDD-Net to other types of tumors, potentially broadening its clinical applicability and impact.