SegAN: Adversarial Network with Multi-scale L₁ Loss for Medical Image Segmentation

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Traditional Generative Adversarial Networks (GANs) face challenges in providing effective feedback for pixel-level tasks.
• Existing methods like U-net may not fully capture complex spatial relationships in medical images.

Purpose of the Study:

• To propose SegAN, an end-to-end adversarial neural network for enhanced medical image segmentation.
• To address the limitations of standard GANs in generating stable gradients for dense prediction tasks.
• To improve the accuracy and stability of medical image segmentation by capturing both global and local features.

Main Methods:

• Developed SegAN, an adversarial network utilizing a fully convolutional neural network as a segmentor.
• Introduced a novel adversarial critic network with a multi-scale L1 loss function.
• Trained the segmentor and critic networks in an alternating min-max game to optimize feature learning.

Main Results:

• SegAN demonstrates superior effectiveness and stability compared to the U-net segmentation method.
• On the BRATS 2013 dataset, SegAN achieved state-of-the-art performance for whole and core tumor segmentation, with improved precision and sensitivity for Gd-enhanced tumor core.
• On the BRATS 2015 dataset, SegAN surpassed state-of-the-art methods in both Dice score and precision.

Conclusions:

• The proposed SegAN framework with a multi-scale loss function is highly effective for medical image segmentation.
• SegAN offers improved performance and stability, particularly for challenging tasks like brain tumor segmentation.
• The novel adversarial critic and multi-scale loss contribute to capturing essential spatial relationships for accurate segmentation.