PEEK Cranioplasty as an Acoustic Window for Transcranial Sonography: Blinded Image Quality Assessment by Clinicians With Exploratory AI Concordance Analysis.

OBJECTIVE

To compare transcranial sonography (TCS) image quality between polyetheretherketone (PEEK) cranioplasty and decompressive craniectomy, and to explore clinician-artificial intelligence (AI) concordance in postoperative neuroimaging interpretation.

METHODS

We retrospectively analyzed TCS images from patients who had undergone PEEK cranioplasty or decompressive craniectomy. All images were preprocessed and blindly scored on a 5-point Likert scale. Evaluations were conducted by clinicians, including both senior and junior doctors, as well as a GPT-4o-based large language model (LLM), with further stratification based on probe frequency. Dataset imbalance was addressed using the Synthetic Minority Oversampling Technique (SMOTE). Difference, equivalence, and concordance analyses were conducted.

RESULTS

A total of 92 images from 23 patients (PEEK cranioplasty: n = 5; decompressive craniectomy: n = 18) were analyzed, generating 896 ratings. The Mann-Whitney U test found no significant difference in image quality between the PEEK cranioplasty and decompressive craniectomy groups (all P > 0.05). Equivalence testing confirmed image quality was within predefined margins. Subgroup analysis by probe frequency showed comparable imaging performance. Consistency analysis showed good agreement between clinician and GPT-4o ratings (ICCs = 0.829, 0.839), supported by Bland-Altman plots. Agreement decreased with low-frequency probes (ICCs = 0.750, 0.735) and dropped further with high-frequency probes (ICCs = 0.445, 0.472).

CONCLUSIONS

PEEK cranioplasty provides an effective acoustic window for TCS, supporting its potential as a postoperative neuro-monitoring tool. AI-based image assessment demonstrates potential as an objective and consistent adjunct to clinician evaluation in neurosonographic practice.