.png){kind=icon}
.png){kind=icon}
Return
A wrapper-based learning framework for papillary thyroid carcinoma diagnosis using optimized feature selection
.png){kind=icon}
PRE
AI
DOI：10.1016/j.bspc.2026.109661.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Papillary thyroid carcinoma (PTC) is the most common subtype of thyroid cancer, comprising roughly 80–90% of all cases. The current diagnosis relies on the subjective interpretation of cytology and histology. The varying experience levels of pathologists lead to different conclusions and interobserver variability. Some pathologists struggle to accurately identify minor or mild lesions, which can result in misjudgments and affect the accuracy and comprehensiveness of the diagnosis. To address the limitations of subjective interpretation and interobserver variability, this study introduces a robust data-driven framework. We propose a wrapper-based machine learning framework that combines a novel hybrid optimizer—Particle Swarm guided Polar Lights Optimizer with Linnik Flight (PS-PLO)—with Kernel Extreme Learning Machines (KELM). Unlike standard metaheuristics, PS-PLO introduces three specific innovations to mitigate premature convergence: (1) integrating PSO-based velocity updates to accelerate global search, (2) employing Linnik flight perturbations with heavy-tailed distributions to escape local optima, and (3) utilizing self-adaptive control mechanisms to balance exploration and exploitation. PS-PLO guides feature selection via velocity-driven updates, auroral oval exploration, and Linnik–flight perturbations, which together maintain a dynamic balance between exploration and exploitation. We first demonstrate PS–PLO’s global search robustness and rapid convergence on the IEEE CEC2017 benchmark suite. Next, we apply the bPSPLO–KELM to a real-world PTC-POP dataset and compare its diagnostic accuracy with recent machine learning baselines. Our method achieves an accuracy of 96.756%, a sensitivity of 96.873%, and a specificity of 97.211% while reducing feature dimensionality through subset selection. This transparent, data-driven decision support tool aims to reduce diagnostic uncertainty and improve patient management in PTC.
Journal
IF:
4.9
Papers:
9.7K
Citations:
2.4W
Organization
View more organizations