Related Experiment Video
Updated: Jun 4, 2025
Author Spotlight: Integrating Ultrasound Imaging with Biochemical Markers for Thyroid Disease Diagnosis
Published on: February 9, 2024
Artificial Intelligence and Whole Slide Imaging Assist in Thyroid Indeterminate Cytology: A Systematic Review
Olia Poursina1, Azadeh Khayyat2, Sara Maleki1
1Department of Pathology and Laboratory Medicine, Brown University, Providence, Rhode Island, USA.
Introduction:
Thyroid cytopathology, particularly in cases of atypia of undetermined significance/follicular lesions of undetermined significance (AUS/FLUS), suffers from suboptimal sensitivity and specificity challenges. Recent advancements in digital pathology and artificial intelligence (AI) hold promise for enhancing diagnostic accuracy. This systematic review included studies that focused on diagnostic accuracy in AUS/FLUS cases using AI, whole slide imaging (WSI), or both.
Methods:
Of the 176 studies from 2000 to 2023, 13 met the inclusion criteria. The datasets range from 145 to 964 WSIs, with an overall number of 494 AUS cases ranging from eight to 254. Five studies used convolutional neural networks (CNNs), and two used artificial neural networks (ANNs). The preparation methods included Romanowsky-stained smears either alone or combined with Papanicolaou-stained or H&E and liquid-based cytology (ThinPrep). The scanner models that were used for scanning the slides varied, including Leica/Aperio, Alyuda Neurointelligence Cupertino, and PANNORAMIC™ Desk Scanner. Classifiers used include Feedforward Neural Networks (FFNNs), Two-Layer Feedforward Neural Networks (2L-FFNNs), Classifier Machine Learning Algorithm (MLA), Visual Geometry Group 11 (VGG11), Gradient Boosting Trees (GBT), Extra Trees Classifier (ETC), YOLOv4, EfficientNetV2-L, Back-Propagation Multi-Layer Perceptron (BP MLP), and MobileNetV2.
Results:
The available studies have shown promising results in differentiating between thyroid lesions, including AUS/FLUS. AI can be especially effective in removing sources of errors such as subjective assessment, variation in staining, and algorithms. CNN has been successful in processing WSI data and identifying diagnostic features with minimal human supervision. ANNs excelled in integrating structured clinical data with image-derived features, particularly when paired with WSI, enhancing diagnostic accuracy for indeterminate thyroid lesions.
Conclusion:
A combined approach using both CNN and ANN can take advantage of their strengths. While AI and WSI integration shows promise in improving diagnostic accuracy and reducing uncertainty in indeterminate thyroid cytology, challenges such as the lack of standardization need to be addressed.
