Explainable Machine Learning Framework for Thyroid Disease Detection: A Hybrid ML–XAI Approach with Clinical and Imaging Data

Abstract

Thyroid problems are one of the most common disorders of the endocrine system. These diagnoses must be accurate and timely in order to be able to treat them and assist professionals to make treatment choices. In this paper we deploy deep learning along with Explainable Artificial Intelligence (XAI) methods to classify thyroid abnormalities based on ultrasound data. In order to deal with the problem of class imbalance, a dataset of 3,538 thyroid ultrasound pictures was partitioned into three classes: benign, malignant, and normal. Segmentation, augmentation and normalization were used to preprocess the data. Both a custom Convolutional Neural Network (CNN) and a pre-trained VGG16 were utilized. Some of the XAI models used to make things simple included SHAP, LIME, Grad-CAM and Integrated Gradients. The experimental results had a good test accuracy of 85.65 and demonstrated that VGG16 always converged both in the training and validation phases, performing better than the custom CNN. Whereas LIME concluded that smooth edges reflected the nodules to be not dangerous, SHAP insisted that the shape, edges and texture of the nodules were important features. Nodule locations of clinical relevance were shown in Grad-CAM output, and the number of nodules that Integrated Gradients agreed with radiologist remarks was high (82), showing a high degree of agreement between professional image analysis. Its results show that XAI can be employed to complement deep learning to offer accurate and understandable diagnostic assistance. Transfer learning with VGG16 and XAI can enhance the accuracy of AI-assisted detection of thyroid illnesses.

Ähnliche Arbeiten