# Advanced AI-Powered System for Comprehensive Thyroid Cancer Detection and Malignancy Risk Assessment

## TL;DR

This paper presents a new AI system that combines ultrasound imaging and molecular data to detect thyroid cancer and assess malignancy risk with high accuracy.

## Contribution

A novel hybrid CAD system integrating CNNs and molecular data for comprehensive thyroid cancer diagnosis and malignancy prediction.

## Key findings

- The CNN module achieved 93.65% accuracy in classifying thyroid nodules from ultrasound images.
- The gene analysis module showed strong performance with low training and testing mean squared error values.

## Abstract

The thyroid cancer incidence has been continuously rising over the last decades. Recently, intelligent cancer detection software are gaining popularity, due to their high diagnostic accuracy and subsequent direct benefits in avoiding unnecessary surgical interventions. This study introduces a novel hybrid computer-aided diagnosis (CAD) system that combines convolutional neural networks (CNNs) and molecular data analysis to achieve comprehensive and reliable thyroid cancer diagnostics. The system consists of two key modules: The first is a CNN-based model leveraging transfer learning, processes ultrasound images to classify patients as either “healthy” or “with a thyroid nodule.” In cases where a nodule is detected, the second module utilizes molecular data to predict the malignancy risk, providing a probability score for clinical decision support. Different image augmentation techniques (traditional ones as well as novels) were carried out to enhance the robustness of the system. The combination of two independent modules makes it possible to use them decoupled, while used together they provide a powerful, in-depth diagnosis of thyroid cancer. The proposed system demonstrates strong performance: the ultrasound-based CNN module achieves an accuracy of 93.65%, with a sensitivity of 100% and a specificity of 69.23%. For the gene analysis component, the model achieves a training mean squared error (MSE) of 4.24 × 10−5 and a testing MSE 6.31 × 10−3. These results underscore the system’s competitive performance with existing thyroid cancer detection CAD systems in both diagnostic performance and the depth of insights provided, supporting clinicians in making informed, reliable decisions in thyroid cancer management.