Enhancing Diagnostic Precision in Thyroid Nodule Classification: A Deep Learning Approach to Automated Ultrasound Image Analysis

Abstract

ABSTRACT Introduction Escalating thyroid nodule prevalence necessitates precise ultrasonographic diagnosis, which is constrained by operator-dependent variability. Convolutional neural network (CNN)-based artificial intelligence (AI)/machine learning (ML) frameworks can improve segmentation, malignancy prediction, and interobserver concordance, yet they often lack real-world clinical validation, interpretable architectures, and actionable validation frameworks for translational integration. Objective To improve diagnostic accuracy in thyroid nodule classification using a deep learning (DL) approach for automated analysis of ultrasound images. Method This methodology employed a multicenter, retrospective cohort of anonymized thyroid ultrasound images (benign/malignant, histopathology-confirmed) sourced from PubMed®. Images were preprocessed (normalization, denoising) with expert-annotated regions of interest (ROIs). A CNN-based DL framework (ResNet-50, EfficientNet-B0) was fine-tuned via transfer learning for automated nodule detection, segmentation, and malignancy classification aligned with ACR TI-RADS™ criteria. Validation utilized an independent test set, diagnostic metrics (sensitivity, specificity, area under the receiver operating characteristic curve (AUC-ROC)), and interobserver analysis (Cohen’s kappa) against three sonographers. Statistical rigor included PSPP-driven paired t-tests, chi-square tests, and McNemar’s tests to quantify AI-human concordance and optimize ACR TI-RADS™ integration for risk stratification. Results The AI model demonstrated high diagnostic efficacy: sensitivity 92.5%, specificity 88.3%, accuracy 90.4%, and AUC-ROC 0.94, surpassing sonographers in both sensitivity (p<0.001) and specificity (p<0.01). Interobserver concordance (Cohen’s κ=0.89) exceeded human variability (κ=0.72–0.85). ACR TI-RADS™ integration achieved 91.2% agreement, enhancing objectivity in the assessment of intermediate-risk nodules (categories 3–4). Feature analysis highlighted robust detection of hypoechoic patterns (94.2% sensitivity) and irregular margins (91.8% sensitivity), aligning with ACR TI-RADS™ criteria and confirming the AI’s potential to standardize risk stratification and reduce diagnostic subjectivity. Conclusion Advanced AI enhances thyroid ultrasound diagnostics through precise nodule detection and classification, reduced interobserver variability, and ACR TI-RADS™-aligned feature extraction, thereby boosting diagnostic confidence and clinical decision-making.

Ähnliche Arbeiten