Deep Learning-Based Automated Diagnosis of Thyroid Nodules from 2D Ultrasound Using a Botox-Optimized Gates-Controlled Deep Unfolding Network

Abstract

Thyroid nodules in clinical practice are usually assessed by experienced doctors based on 2D ultrasound images to decide if fine needle aspiration (FNA) is required. This subjective evaluation, however, results in unnecessary FNAs, subjecting patients to discomfort and additional medical expenses. This research proposes the creation of an automatic computer-aided diagnostic system for thyroid nodule classification to eliminate unnecessary FNAs. The study suggests a system based on deep learning that combines cutting-edge image preprocessing, feature extraction, and classification. Adaptive and Propagated Mesh Filtering (APMF) is first applied to improve image quality by eliminating noise and edge details in ultrasound images. Subsequently, discriminative features are extracted by Two-Sided Clifford Wave Transforms (TSCWT), which pick up both local and directional information from the images efficiently. For classification, the study uses a Gates-Controlled Deep Unfolding Network (GCDUN) optimized by the Botox Optimization Algorithm (BOA) to achieve efficient learning with better convergence and resilience. The dataset is 591 thyroid ultrasound images labeled under the Bethesda scoring system, categorized into two classes: nodules that need to be subjected to FNA and nodules that don't need to be subjected to FNA. The system proposed in this work had great performance, with an average accuracy of 99.86%, an F1 score of 99.81%, and a specificity of 99.84%. The system provides high diagnostic performance and low false negatives, with strong potential to minimize unnecessary FNAs in clinical practice.

Ähnliche Arbeiten