RiIG-Based Contourlet Parametric Imaging and Reliability-Guided Learning for Breast Ultrasound Lesion Classification

Abstract

Breast ultrasound (BUS) lesion classification remains challenging because of speckle noise, low contrast, tissue heterogeneity, and acquisition variability. This study presents a framework combining Rician Inverse Gaussian (RiIG)-based contourlet parametric imaging with reliability-guided learning. The method applies contourlet decomposition and RiIG statistical modeling to transform B-mode images into parametric representations of multiscale directional structure and heterogeneous lesion echotexture. A reliability-weighted DenseNet–Vision Transformer fusion network and teacher–student procedure are optimized using a structured composite objective, hereafter termed LoFT, which coordinates hard-sample learning, imbalance-sensitive supervision, confidence control, and perturbation consistency using established loss terms rather than introducing a new loss family. The fusion mechanism adaptively integrates local echotextural and lesion-boundary evidence from DenseNet with global contextual information from the Transformer. Evaluation included leakage-aware five-fold experiments on the publicly available Breast Ultrasound Images (BUSI) dataset, UDIAT Diagnostic Centre dataset (UDIAT), Mendeley dataset, and their combined corpus; strict BUSI-to-UDIAT transfer using the BUSI-only model and threshold selection; and supplementary source-combined unseen-target testing. Proxy-group-disjoint and duplicate-audited partitions were used, although verified patient identifiers were unavailable. Five-fold accuracies were 95.58%, 86.44%, 99.60%, and 92.21%, respectively. Strict external transfer achieved 76.07% accuracy and 77.56% area under the receiver operating characteristic curve (AUC); however, 37.04% malignant sensitivity and 50.63% malignant-class F1-score indicated substantial domain-shift limitations. Supplementary evaluations further revealed inconsistent transfer across unseen target datasets. Gradient-weighted Class Activation Mapping (Grad-CAM), Integrated Gradients, Partial Dependence Plots and quantitative BUSI mask analysis showed partial correspondence between model attribution and annotated lesion regions. The results demonstrate strong within-dataset technical performance while exposing unresolved external-generalization limitations. The framework therefore supports further retrospective decision-support research while underscoring the need for improved cross-dataset robustness and prospective clinical validation.