A comparative analysis of deep learning architectures for thyroid tissue classification with hyperspectral imaging

Abstract

Hyperspectral imaging has shown significant applicability in the medical field, particularly for its ability to represent spectral information that can differentiate specific biomolecular characteristics in tissue samples. However, the complexity of analyzing HSI data, due to its high dimensionality and the large volume of information, presents significant challenges. At the same time, deep learning, particularly convolutional neural networks and recurrent neural networks, has become an essential tool in medical diagnostics, providing detailed analysis across various contexts. These techniques enable the analysis of complex information often unattainable through traditional methods. This paper introduces a novel approach that integrates micro-FTIR spectroscopy with three different deep learning architectures, namely RNN, FCNN, and 1D-CNN, to compare their performance in region-based classification of thyroid tissues, including goiter, cancerous, and healthy types. The proposed deep learning methods were developed on a dataset of 60 patients and evaluated using grouped 10-fold cross-validation. The 1D-CNN achieved the highest scores in classifying the spectral data provided by micro-FTIR, enabling more precise and accurate region-based tissue classification. The 1D-CNN achieved an accuracy of 97.60%, while RNN and FCNN achieved 96.88% and 93.66%, respectively. These results highlight the effectiveness of this approach in enhancing the precision of thyroid pathology analysis.