Multispectral Imaging as an Adjunct for Classification of Thyroid Fine Needle Aspirates.
Lewis D Hahn, David Rimm, Constantine Theoharis. Yale University School of Medicine, New Haven, CT
Background: Fine-needle aspiration (FNA) biopsy is the standard method for evaluation of thyroid nodules. The Bethesda System improves the reliability of FNA diagnosis through standardization of diagnostic features, but equivocal cases persist. Multispectral imaging of FNA biopsies captures visible light information beyond the capacity of the human eye and could increase the accuracy of FNA biopsy in classifying challenging cases. In this study, we focus on the distinction between papillary carcinoma (PTC) and goiter, and the distinction between follicular carcinoma (FC) and follicular adenoma (FA).
Design: 84 archived cytology cases utilizing the Bethesda System from 2007-2009 were collected. A CRI Nuance multispectral camera and the accompanying software were used to image the thyroid FNAs and develop computer-based classification algorithms (classifiers). We developed two classifiers. The first, "PTC/G," classifies FNA images as either PTC or goiter. The second, "FC/FA," classifies FNA images as either FC or FA. Classifiers segregate images into regions based on represented wavelengths and then compute areas of the regions; in the case of PTC/G, this corresponds to regions of PTC, goiter, or neither feature. If the ratio of PTC area to goiter area is over a threshold value, the image as a whole is classified as PTC; FC/FA functions similarly. PTC/G was developed using 40 images of PTCs and goiters and tested on a distinct set of 30 PTC and 30 goiter images. FC/FA was developed using 35 images taken from surgically confirmed FC and FA cases and tested on a distinct set of 14 FC and 15 FA images.
Results: A Receiver Operating Characteristic (ROC) curve was generated for the PTC/G classifier with an Area Under the Curve (AUC) of 0.90. The AUC of the FC/FA classifier ROC curve was 0.76. Choosing threshold ratios of malignant to benign areas for each classifier, we generated three specific tests which are summarized in the table.