Authors:
Ahana Choudhury
1
;
Radu Mihail
1
and
Sorin Chiriac
2
Affiliations:
1
Department of Computer Science, Valdosta State University, Valdosta, U.S.A.
;
2
University of Medicine and Pharmacy “Victor Babes” Timisoara, Romania
Keyword(s):
Thyroid Cancer, Thyroid Nodule, Classification, Convolutional Neural Networks, Explainable AI, Interpretability.
Abstract:
The gold standard in thyroid nodule malignancy diagnosis consists of ultrasound (US or sonogram) guided fine needle aspiration biopsy. This procedure is ordered based on an assessment of malignancy risk by a trained radiologist, who uses US images and relies on experience and heuristics that are difficult to effectively systematize into a working algorithm. Artificial Intelligence (AI) methods for malignancy detection in sonograms are designed to either perform segmentation (highlight entire thyroid gland and/or nodule) or output a probability of malignancy. There is a gap between AI methods trained to perform a specific task using a black-box method, and the sonogram features (e.g.,: shape, size, echogenicity, echotexture) that a radiologist looks at. We aim to bridge this gap, using AI to reveal saliency in sonograms for features that are easily understood by clinicians. We propose a deep-learning model that performs two tasks important to radiologists: sonogram feature saliency de
tection, as well as probability of malignancy. We perform both a quantitative and qualitative evaluation of our method using an open dataset, the Thyroid Digital Image Database (TDID). Our framework achieves 72% accuracy in the task of classifying thyroid nodules as benign or malignant.
(More)