AI-Based Clinical Decision Support System for Thyroid Nodule Classification in Ultrasound Imaging with Web-Based Implementation

Abstract

Thyroid nodules are a frequent clinical phenomenon and proper classification into the normal, benign, and malignant types is required to diagnose and plan treatment at an early phase. One of the most common imaging techniques is ultrasound, which is non-invasive, economical, but its interpretation is subjective, and it requires the expertise of the radiologist. In this study, a novel intelligent artificial intelligence (AI) system of diagnostic thyroid nodules through ultrasound imaging is offered. An efficient preprocessing pipeline, comprising of image resizing, intensity normalization, CLAHE-based contrast enhancement, median filtering, and data augmentation, was used to improve image quality and the performance of the model. A variety of deep learning models were tested (a baseline convolutional neural network (CNN)) and various transfer learning architectures. The CNN baseline obtained 76% accuracy and transfer learning models greatly enhanced performance. VGG16 and MobileNetV2 were capable of 84% and 88% accuracy, respectively, but ResNet50 scored 91 %. EfficientNet achieved the highest performance of 94% accuracy, high recall, precision, and F1-score. The findings show that transfer learning was effective in classifying thyroid ultrasound and the proposed system has potential to be an effective clinical decision-support tool in enhancing diagnostic accuracy and early diagnosis of malignant cases.