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

Ultrasound image segmentation of diffuse liver fibrosis nodules confronts three critical challenges: boundary ambiguity caused by gradual tissue transitions, texture heterogeneity arising from fibrotic variations, and inadequate uncertainty quantification, which manifests as overconfident misclassifications at fibrotic nodules.

To address these challenges, this paper proposes the Edge-Semantics Probabilistic Dirichlet Network (ESPD-Net), which integrates Dirichlet evidential theory into diffuse lesion segmentation for nodule detection. ESPD-Net employs three synergistic components: (1) The Semantic-Probabilistic Dual Path Fusion (SPDF) bottleneck constructs parallel semantic and probabilistic pathways to capture local morphological and global distribution features. (2) The Dirichlet Evidential Guided Decoder (DEGD) reformulates segmentation as second-order probabilistic modeling under evidential theory, guiding the adaptive feature decoding process by outputting calibrated uncertainty distributions. (3) Guided by these localized high-uncertainty regions, the Dirichlet Boundary Aware Refinement (DBAR) module executes targeted geometric corrections to precisely align ambiguous lesion boundaries.

Evaluations on murine and clinical datasets demonstrate that ESPD-Net significantly outperforms state-of-the-art methods. Specifically, it achieves a Dice of 0.855 (+0.043) and an IoU of 0.747 (+0.057). Furthermore, the model effectively minimizes calibration and boundary errors, reducing ECE to 3.85% and HD95 to 3.25.

These findings demonstrate that the proposed ESPD-Net effectively addresses the challenges of diffuse lesion segmentation, thereby objectively confirming its clinical potential for computer-aided diagnosis.