HUR-MACL: High-Uncertainty Region-Guided Multi-Architecture Collaborative Learning for Head and Neck Multi-Organ Segmentation

TL;DR

This model adaptively identifies high uncertainty regions using a convolutional neural network, and for these regions, Vision Mamba as well as Deformable CNN are utilized to jointly improve their segmentation accuracy.

Abstract

Accurate segmentation of organs at risk in the head and neck is essential for radiation therapy, yet deep learning models often fail on small, complexly shaped organs. While hybrid architectures that combine different models show promise, they typically just concatenate features without exploiting the unique strengths of each component. This results in functional overlap and limited segmentation accuracy. To address these issues, we propose a high uncertainty region-guided multi-architecture collaborative learning (HUR-MACL) model for multi-organ segmentation in the head and neck. This model adaptively identifies high uncertainty regions using a convolutional neural network, and for these regions, Vision Mamba as well as Deformable CNN are utilized to jointly improve their segmentation accuracy. Additionally, a heterogeneous feature distillation loss was proposed to promote collaborative learning between the two architectures in high uncertainty regions to further enhance performance. Our method achieves SOTA results on two public datasets and one private dataset.

Figures (9)

• Figure 1: (A). Schematic diagram of HNOARs. BS (Brainstem), RE (Right Eye), LE (Left Eye), RON (Right Optic Nerve), LON (Left Optic Nerve), RPG (Right Parotid Gland), LPG (Left Parotid Gland), RT (Right Temporal lobe), LT (Left Temporal lobe), RM (Right Mandible), LM (Left Mandible). (B). Segmentation error cases, where white denotes the ground truth and color denotes the predictions. (C). Segmentation architectures, where different colors represent different networks.
• Figure 2: The overall framework of HUR-MACL, which involves hard region mining to mask non-hard regions, followed by multi-architecture collaborative segmentation (using ViM and DCNN) on the masked feature map, and heterogeneous feature distillation module is introduced to exchange reliable information in hard regions.
• Figure 3: The overall framework of Vision Mamba Encoder.
• Figure 4: Visualizations on the PDDCA dataset, where red dashed boxes indicate magnified regions and blue dashed boxes indicate non-magnified regions.
• Figure 5: Visualizations on the Structseg dataset, where red dashed boxes indicate magnified regions and blue dashed boxes indicate non-magnified regions.