External validation of a deep learning-based algorithm for detection of tall cells in papillary thyroid carcinoma: A...

Area of Science:

• Oncology
• Pathology
• Artificial Intelligence in Medicine

Background:

• Tall cell papillary thyroid carcinoma (TC-PTC) is an aggressive subtype of papillary thyroid carcinoma (PTC).
• The current definition of TC-PTC relies on subjective morphological criteria, leading to significant inter-observer variability in diagnosis.
• Accurate identification of TC-PTC is crucial for prognostication and treatment planning.

Purpose of the Study:

• To validate a deep learning (DL)-based algorithm for the objective detection of tall cells in papillary thyroid carcinoma.
• To assess the performance of the DL algorithm on externally collected, whole-slide images of TC-PTC.
• To evaluate the correlation between TC-PTC identification by the DL algorithm and patient relapse-free survival.

Main Methods:

• A previously trained DL algorithm was applied to 160 externally collected hematoxylin and eosin (HE)-stained PTC whole-slide images.
• The DL algorithm's performance was evaluated using 360 manual annotations from 18 tissue sections.
• Sensitivity and specificity were calculated for the detection of tall cells (TCs) and non-TC areas.

Main Results:

• The DL algorithm demonstrated high sensitivity (90.6%) and specificity (88.5%) in detecting TCs.
• The algorithm showed high sensitivity (81.6%) and specificity (92.9%) in identifying non-TC areas.
• TC thresholds of 20% and 30% identified by the DL algorithm significantly correlated with shorter relapse-free survival.

Conclusions:

• The DL-based algorithm reliably detects TCs in unseen, external HE-stained PTC slides without retraining.
• This AI tool offers an objective and reproducible method for TC-PTC identification.
• The algorithm's findings correlate with clinical outcomes, supporting its potential utility in improving TC-PTC diagnosis and patient management.