Unsupervised Classification of THYROID Cancer using Multiwavelength Photoacoustic Imaging

The incidence of thyroid cancer has been increasing in the developing countries in the last few years. Although currently, Ultrasound imaging is the gold standard for detecting thyroid malignancy, it suffers from several limitations. Photoacoustic imaging, which provides the functional information of the tissue in terms of spatially varying optical properties, can potentially be used for early detection of thyroid cancer. In this retrospective study, we performed unsupervised classification for the detection of thyroid malignancy with mutiwave-length photoacoustic data, generated by freshly excised thyroid specimens from actual human patients. Autoencoder and sparse auto-encoder networks were specifically tuned for the classification of malignant and nonmalignant thyroid. The performance of these networks was compared with that of standard unsupervised classifiers, K-means and Gaussian Mixture Model algorithms. The preliminary results show that it is possible to perform unsupervised classification with moderate accuracy for thyroid cancer detection using photoacoustic imaging. The sparse auto-encoder obtained the highest clustering accuracy among all the four classifiers. It was also observed that the network based classifiers performed better than K-means and Gaussian Mixture Model algorithms.