Data-Driven Deep Supervision for Medical Image Segmentation.

Area of Science:

• Medical image analysis
• Computer vision
• Deep learning

Background:

• Medical image segmentation is crucial for diagnosis but faces challenges like low contrast and complex objects.
• Existing methods struggle with robust feature extraction due to data-specific complexities.
• Difficulties in dense prediction hinder accurate segmentation outcomes.

Purpose of the Study:

• To develop a robust deep convolutional neural network (CNN) for improved medical image segmentation.
• To enhance feature extraction by exploiting data-specific attributes for better dense prediction.
• To address challenges in medical image segmentation caused by low contrast, noise, and complex object boundaries.

Main Methods:

• Proposed a novel deep convolutional neural network (CNN) architecture.
• Introduced a method for matching object perceptive field (OPF) with layer-wise effective receptive fields (LERF) for strategic deep supervision.
• Developed densely decoded networks (DDN) with additional connections for refined dense prediction and localization.

Main Results:

• The proposed CNN effectively extracts robust, data-specific features.
• Strategic deep supervision using OPF-LERF matching improved feature extraction.
• Densely decoded networks (DDN) achieved better target localization and refined dense prediction.
• Validated effectiveness across diverse 2D and 3D datasets including retinal vessels, melanoma, optic disc/cup, spleen, lymph node, and fungus segmentation.

Conclusions:

• The proposed CNN approach significantly enhances medical image segmentation accuracy.
• The novel deep supervision strategy and DDN architecture improve feature extraction and prediction.
• The method demonstrates broad applicability and effectiveness across various medical imaging segmentation tasks.