A novel diffuse liver nodule detector via integrating semantic edge features and probabilistic uncertainty modeling

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Ultrasound image segmentation of diffuse liver fibrosis nodules faces challenges including boundary ambiguity, texture heterogeneity, and poor uncertainty quantification leading to misclassifications.
• Accurate segmentation is crucial for diagnosis and treatment planning of liver fibrosis.

Purpose of the Study:

• To introduce the Edge-Semantics Probabilistic Dirichlet Network (ESPD-Net) for improved diffuse liver fibrosis nodule segmentation.
• To address boundary ambiguity and uncertainty quantification in ultrasound liver fibrosis imaging.

Main Methods:

• The ESPD-Net integrates Dirichlet evidential theory with a dual-path fusion bottleneck (SPDF) for parallel semantic and probabilistic pathway analysis.
• A Dirichlet Evidential Guided Decoder (DEGD) reformulates segmentation as second-order probabilistic modeling, outputting calibrated uncertainty distributions.
• A Dirichlet Boundary Aware Refinement (DBAR) module uses high-uncertainty regions to correct ambiguous lesion boundaries.

Main Results:

• ESPD-Net significantly outperforms state-of-the-art methods on murine and clinical datasets, achieving a Dice score of 0.855 and IoU of 0.747.
• The model effectively minimizes calibration errors (ECE to 3.85%) and boundary errors (HD95 to 3.25).

Conclusions:

• The proposed ESPD-Net successfully addresses key challenges in diffuse liver fibrosis segmentation.
• The method demonstrates significant potential for clinical application in computer-aided diagnosis of liver fibrosis.