Sitemap

Thyroid nodule malignancy assessment is a high-stakes predictive task in clinical radiology, where diagnostic uncertainty directly affects biopsy referral, surgical planning, and patient outcomes. We present SSMURNet, a deep learning framework for confidence-calibrated thyroid nodule classification from ultrasound images. The architecture integrates a dual-branch spectral–spatial encoder combining FFT frequency analysis with Mamba state space modelling, an Uncertainty-Guided Attention (UGA) mechanism, and a hierarchical evidential learning objective. Evaluated on 7,288 pathologically confirmed images, SSMURNet achieved 92.50% accuracy and 0.9700 AUROC. External validation on an independent dataset achieved 85.42% accuracy and 0.9185 AUROC without retraining, demonstrating robust cross-domain generalisation.

Breast cancer early detection heavily relies on ultrasound imaging, but accurate diagnosis is often hindered by speckle noise, operator dependency, and indistinct lesion boundaries. Existing deep learning methods lack hierarchical multi-scale integration, quantitative interpretability validation, and cross-dataset generalization analysis, hindering clinical adoption. In this work, we propose HyFormer-Net, a hybrid CNN-Transformer framework integrating EfficientNet-B3 and Swin Transformer via multi-scale hierarchical fusion blocks at four encoder stages. The attention-gated decoder enables dual-pipeline interpretability: intrinsic attention validation (quantitative IoU verification) and Grad-CAM classification reasoning. HyFormer-Net achieved 76.1% Dice score and 93.2% classification accuracy on BUSI, with clinically critical 92.1% Malignant Recall.

Deep learning methods dominate thyroid nodule segmentation in ultrasound images, achieving high Dice scores but at the cost of millions of parameters, GPU-dependent training via backpropagation, and limited mathematical tractability. We propose MedSaab-US, a backpropagation-free segmentation framework grounded in the Green Learning paradigm. MedSaab-US extracts multi-scale spatial-frequency features by combining multi-level Discrete Wavelet Transform (DWT) with multi-scale channel-wise Saab transforms at patch sizes of 5×5, 11×11, and 21×21 pixels. Label-Assisted Greedy (LAG) feature selection retains the most discriminative features, which are fed to an XGBoost classifier for pixel-wise prediction. Evaluated on the TN3K dataset (2,879 training and 614 test images), MedSaab-US achieves a mean Dice coefficient of 0.4784 ± 0.2190 with a model footprint under 500K parameters and CPU-only inference in approximately 0.3 seconds per image.

Deep learning has achieved remarkable performance in medical image segmentation, yet it suffers from critical limitations: mathematical intractability, substantial parameter requirements, and lack of clinical interpretability. We propose RadiomicNet, a novel two-stream hybrid architecture that enhances standard deep learning by integrating handcrafted radiomics features directly into the segmentation learning process. The key contribution is the Radiomics Attention Gate (RAG), which leverages Gray-Level Co-occurrence Matrix (GLCM) and Local Binary Pattern (LBP) features to modulate skip-connection attention in a lightweight MobileNetV2-based encoder-decoder, providing ante-hoc interpretability without post-hoc approximations. RadiomicNet achieves a DSC of 0.763 ± 0.231 on BUSI and 0.854 ± 0.112 on Kvasir-SEG, outperforming U-KAN by 1.2% and 1.8% respectively, with only 3.27M parameters.