Improving Thyroid Cancer Diagnosis and Risk Stratification with VGGNet-Driven Deep Learning from Multimodal Imaging

Abstract

Recent developments in deep learning have demonstrated potential in improving the analysis of medical images in diagnosing cancer. This paper tackles the problem of precise diagnosis of thyroid cancer and stratification of risk based on multimodal imaging data. We suggest a deep-learning structure using the VGGNet framework, which utilizes ultrasound and other imaging modalities, to automatically discover and combine discriminative features that are pertinent to thyroid pathology. The model was trained and tested on a large, labelled dataset and its performance was measured by metrics, including accuracy, ROC-AUC, confusion matrices and probability histograms of prediction. Findings indicate that the VGGNet-based model had good training and validation accuracy, ROC-AUC scores of up to 0.98 and strong classification accuracy in various thyroid cancer subtypes. Prediction probabilities were also well-calibrated to support the model with clear separations in the confusion matrix. We find that multimodal imaging through VGGNet based and deep learning can substantially enhance diagnostic accuracy of thyroid cancer and individual stratification of risk to aid clinical judgments.