Comparison of texture-based classification and deep learning for plantar soft tissue histology segmentation

Area of Science:

• Biomedical Engineering
• Digital Pathology
• Computational Anatomy

Background:

• Histomorphological measurements are crucial for identifying disease-related microstructural changes, especially in diabetic plantar soft tissues.
• Current measurement methods are time-consuming and prone to significant human and sampling errors.
• Automating segmentation is key to enabling efficient and reliable morphological analysis.

Purpose of the Study:

• To develop and compare automated segmentation approaches for plantar soft tissue.
• To evaluate the performance of a modified U-Net convolutional neural network against traditional feature-based methods.
• To facilitate accurate morphological analysis of diabetic-related plantar soft tissue changes.

Main Methods:

• Investigated two automated segmentation approaches: 1) texture- and color-based features with tile-wise classification, and 2) a modified U-Net convolutional neural network.
• Utilized modified Hart's stain for elastin in plantar soft tissue samples.
• Compared segmentation accuracy, focusing on error types and robustness to variations.

Main Results:

• The modified U-Net achieved superior segmentation performance compared to the best feature-based method (f-IOCLBP), outperforming it by 3.6%.
• Feature-based methods showed sensitivity to illumination and staining intensity variations, with errors in large regions or at tissue boundaries.
• The U-Net effectively segmented small, few-pixel boundaries, with errors easily correctable via post-processing.

Conclusions:

• A modified U-Net deep learning approach significantly outperforms traditional hand-crafted feature methods for segmenting plantar soft tissue stained for elastin.
• Automated segmentation using U-Net offers a more accurate, robust, and efficient solution for histomorphological analysis in conditions like diabetic foot disease.