Hybrid Classification Approach of Thyroid Cancer Histological Images using Deep Feature Extraction and Classical Classifiers

Thyroid tumors present a common clinical challenge, often requiring histological analysis for accurate diagnosis. In this study, we performed the classification of thyroid tumors using a real database of histological images. We studied the efficiency of the classification of three pathologies (medullary carcinoma, papillary carcinoma, and vesicular carcinoma) by hybrid approach combining deep feature extraction with a pre-trained Convolutional Neural Network model (with the choice of EfficientNetB0 or MobileNet) and the final classification by a classic model (with the particular choice of Support Vector Machine or Random Forest). For this, different combinations are explored, through different metrics of performance. Each model provided very satisfactory results, the best value being that of the EfficientNetB0&SVM combination with an overall accuracy of 99%. The results suggest that such a proposed hybrid approach effectively captured the discriminative features of the studied cancers of the thyroid, making it a potential high-performance computer-assisted diagnosis in thyroid pathologies.