GREnet: Gradually REcurrent Network With Curriculum Learning for 2-D Medical Image Segmentation.

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Medical image segmentation is crucial for diagnosis and treatment planning.
• Current deep learning methods for 2-D segmentation face challenges with ambiguity and distractors in ground truth supervision.
• Convolutional Neural Networks (CNNs) have advanced segmentation but direct supervision remains suboptimal.

Purpose of the Study:

• To develop a novel deep learning framework, GREnet, for improved 2-D medical image segmentation.
• To address the limitations of direct ground truth supervision by introducing a curriculum learning approach.
• To formulate 2-D medical image segmentation as a temporal task with pixel-level curriculum learning.

Main Methods:

• Proposed a two-network architecture: GREnet (segmentation) and a curriculum-mining network.
• GREnet formulates segmentation as a temporal task using pixel-level gradual curricula.
• The curriculum-mining network generates curricula of increasing difficulty in a data-driven manner.
• Employed UNet as the backbone for GREnet with ConvLSTM for temporal linking.
• Utilized UNet++ with transformer for the curriculum-mining network.

Main Results:

• Demonstrated the effectiveness of GREnet across seven diverse medical imaging datasets.
• Datasets included dermoscopic, retinal, ultrasound, and CT images.
• Showcased successful segmentation of lesions, optic disc/cup, blood vessels, and lungs.
• The proposed pixel-level curriculum learning approach improved segmentation performance.

Conclusions:

• GREnet offers a significant advancement in 2-D medical image segmentation.
• The novel temporal task formulation with pixel-level curriculum learning effectively handles segmentation challenges.
• This approach shows broad applicability across various medical imaging modalities and segmentation tasks.