From structural complexity to causal representation: A dynamic fractal-attention framework for fine-grained ovarian...

Accurate fine-grained classification of ovarian tumors from ultrasound images remains challenging due to speckle noise, boundary ambiguity, structural heterogeneity, and acquisition-induced spurious correlations. To address these issues, we propose a structure-aware and causally regularized deep learning framework that integrates dynamically learnable fractal modeling, edge-guided structural refinement, and a causal inference-assisted module within a unified architecture. The framework explicitly models lesion-intrinsic structural complexity while suppressing non-discriminative acquisition cues, enabling robust and interpretable subtype recognition. Extensive experiments on a large private ovarian ultrasound cohort and the public MMOTU benchmark demonstrate that the proposed method outperforms representative CNNs, Transformer-based models, self-supervised approaches, and established causal baselines, achieving higher accuracy and macro-F1 with consistently narrower confidence intervals. Patient-level evaluation further shows that dynamic fractal modeling captures disease-intrinsic structural complexity with cross-view stability, leading to substantial improvements in recall and F1-score for structurally complex and low-prevalence subtypes. Comprehensive interpretability analyses, including quantitative structural saliency assessment and Grad-CAM visualization, confirm that predictions are primarily driven by clinically meaningful intralesional structures rather than boundary contrast or acquisition artifacts. A retrospective reader study further demonstrates that AI assistance improves diagnostic consistency and reduces systematic misclassification, particularly among less experienced sonographers. Overall, this work establishes a theoretically grounded and empirically validated framework for consistent and interpretable fine-grained ovarian tumor classification from ultrasound imaging.