Scale-Aware Super-Resolution Network With Dual Affinity Learning for Lesion Segmentation From Medical Images

Area of Science:

• Medical Image Analysis
• Deep Learning
• Computer Vision

Background:

• Convolutional neural networks (CNNs) excel in medical image segmentation but struggle with lesions of varying scales.
• Precise segmentation of small lesions is difficult in low-resolution images, while large lesions require extensive receptive fields.

Purpose of the Study:

• To develop a scale-aware super-resolution (SR) network for adaptive lesion segmentation from low-resolution (LR) medical images.
• To address the challenges posed by diverse lesion sizes and resolutions in medical image segmentation.

Main Methods:

• Proposed a dual-branch network for simultaneous lesion mask SR (LMSR) and lesion image SR (LISR).
• Introduced scale-aware dilated convolution (SDC) blocks in decoders to adapt receptive fields based on lesion size.
• Incorporated feature affinity (FA) and scale affinity (SA) modules to enhance multitask learning between dual branches.

Main Results:

• The proposed network demonstrated consistent improvements over state-of-the-art methods on challenging lesion segmentation datasets.
• Achieved superior performance in segmenting lesions across a wide range of scales and resolutions.

Conclusions:

• The scale-aware super-resolution network effectively segments lesions of various sizes from LR medical images.
• The dual-branch architecture with SDC, FA, and SA modules enhances multitask learning and segmentation accuracy.