Self-supervised Semantic Segmentation: Consistency over Transformation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Supervised deep learning for medical image segmentation requires extensive labeled data, posing a significant challenge.
• Existing methods struggle with data scarcity and accurately delineating complex anatomical structures and lesions.

Purpose of the Study:

• To develop a novel self-supervised algorithm for accurate medical image segmentation.
• To overcome the limitations of data-dependent supervised learning approaches.
• To improve the delineation of lesions with deformations and enhance segmentation robustness.

Main Methods:

• Proposed a self-supervised algorithm integrating Inception Large Kernel Attention (I-LKA) modules for comprehensive contextual information capture.
• Incorporated deformable convolution to effectively handle lesion deformations and improve boundary definition.
• Employed a self-supervised strategy emphasizing invariance to affine transformations and introduced a spatial consistency loss term.

Main Results:

• The algorithm achieved superior performance in skin lesion and lung organ segmentation tasks compared to state-of-the-art methods.
• Demonstrated effective capture of contextual information and preservation of local image intricacies.
• Showcased enhanced ability to model and handle geometric distortions and lesion deformations.

Conclusions:

• The proposed self-supervised method significantly advances medical image segmentation accuracy.
• The algorithm offers a robust solution for segmentation tasks with limited labeled data.
• The integration of I-LKA, deformable convolution, and spatial consistency loss yields state-of-the-art results.