Subject to change.
Subject to change.
Dr. Satturwar is an Associate Professor in the Anatomic Pathology Division and a faculty member of the division of Genitourinary Pathology, Cytopathology and Bone and Soft tissue Pathology at the Ohio State University. Her research interests include understanding pathologic basis of carcinogenesis, exploring novel biomarkers for prostate, kidney, bladder & testicular neoplasms, non-exfoliative cytology and exploring innovative applications of digital pathology and AI.
Background: AIxTHY is an artificial intelligence platform designed to analyze whole-slide images (WSIs) and support cytopathologists in interpreting thyroid fine-needle aspiration cytology (FNAC) within The Bethesda System for Reporting Thyroid Cytopathology (TBS). Multi-Z-layer scanning may overcome the diagnostic limitations of single-layer digital cytopathology. This study evaluated AIxTHY’s performance on single-layer versus 7-Z-layer WSIs, using conventional microscopy as the benchmark, for thyroid FNAC interpretation in routine clinical practice.
Design: 100 ThinPrep FNAC slides with consensus cytologic diagnoses were selected (5 TBS-I, 35 TBS-II, 15 TBS-III, 15 TBS-IV, and 30 TBS-VI cases). Each slide was digitized using a 3DHISTECH scanner to produce paired single-layer (S-WSI) and 7-Z-layer (7-WSI) images. AIxTHY pre-analyzed all WSIs to flag atypical cells for reviewer evaluation. Five reviewers (2 cytopathologists and 3 cytotechnologists) assessed each case across three modalities with two-week washout periods: microscopy (Arm 1), AIxTHY-assisted S-WSIs (Arm 2), and AIxTHY-assisted 7-WSIs (Arm 3). Binary diagnostic accuracy (positive: TBS-III and above; negative: TBS-II), TBS category concordance, and diagnostic turnaround time were compared across 500 total reads.
Results: Sensitivity was higher with AIxTHY (Arm 2: 81.3%; Arm 3: 83.0%) than with microscopy (Arm 1: 68.7%; p<0.001), with no significant difference between Arms 2 and 3 (p=0.522). Specificity was slightly lower with AIxTHY (Arm 2: 68.0%; Arm 3: 65.7%) versus microscopy (Arm 1: 73.1%), reaching significance for Arm 1 versus Arm 3 (p=0.049). Overall accuracy improved with AIxTHY (Arm 2: 76.4%; Arm 3: 76.6%) compared with microscopy (70.3%; p=0.004). Diagnostic efficiency increased substantially, with mean review time reduced by 32.8% (Arm 1: 163.6 sec vs. Arm 2: 109.9 sec; p<0.001). TBS category concordance with consensus was comparable (Arm 1: 52.4%; Arm 2: 51.4%; Arm 3: 53.5%), but agreement for indeterminate TBS-III was higher with AIxTHY (Arm 2: 53.3%; Arm 3: 44.0%) than microscopy (25.3%). Conversely, agreement for benign TBS-II was slightly lower with AIxTHY (Arm 2: 68.0%; Arm 3: 65.7%) than microscopy (73.1%). Nondiagnostic (TBS-I) calls decreased from 10.8% (54/500) with microscopy to 6.4% (Arm 2) and 5.6% (Arm 3), representing a 41–48% relative reduction. Of Arm 1 TBS-I reads, Arms 2 and 3 reclassified 9 and 7 cases to TBS-II, and 4 and 4 cases to ≥TBS-III, respectively.
Conclusion: AIxTHY increased sensitivity, reduced review time, and improved agreement for indeterminate (TBS-III) FNAC, with a modest decrease in specificity. It lowered nondiagnostic rates and enhanced overall diagnostic accuracy compared with conventional microscopy. These findings support the potential of AI-assisted digital cytopathology to streamline thyroid FNAC workflows and improve diagnostic efficiency.