Human-in-the-Loop Enhances Machine Learning Inference in Intraoperative Optical Coherence Tomography Glioma Imaging

Background/Objectives: The integration of Artificial Intelligence (AI) into clinical workflows raises critical questions regarding decision-making responsibility, as fully autonomous systems inevitably carry a margin of error that can be fatal in high-stakes fields like surgery. This study addresses this challenge by evaluating a "Human-in-the-Loop" (HITL) workflow, using intraoperative Optical Coherence Tomography (OCT) for glioma detection. We aimed to determine if integrating Machine Learning (ML)-generated segmentation maps with human contextual analysis resolves the tension between automation and clinical responsibility, yielding superior diagnostic reliability compared to structural or quantitative imaging alone. Methods: We retrospectively analyzed 86 intraoperative OCT scans from 27 patients. Five neurosurgeons blindly assessed the data across three progressive levels of processing: (1) structural scans, (2) physics-based parametric maps, and (3) SVM-based generated segmentation maps. Crucially, the HITL inference performance on segmentation maps was benchmarked against "models-only" inference pipeline: a SVM and a state-of-the-art multimodal reasoning model, Gemini 3.1 Pro. To evaluate interpretability and the operator's ability to confidently exercise their authority, we measured inter-rater consistency alongside diagnostic performance. Results: The results demonstrate that, while quantitative parametric maps improved Global Accuracy (87% [95% CI: 82-92%]) compared to structural scans (80% [95% CI: 73-86%]), they suffered from an "interpretability gap," resulting in a moderate inter-rater consistency of 0.68 [95% CI: 0.59-0.78]. In contrast, the HITL approach using segmentation maps maximized consensus to 0.98 [95% CI: 0.95-1.00] and achieved the highest performance (Accuracy 94% [95% CI: 88-98%] and Sensitivity 98% [95% CI: 92-100%]). Compared to the standalone models, the HITL approach significantly outperformed the SVM baseline (Accuracy 84% [95% CI: 81-87%]; Sensitivity 83% [95% CI: 78-88%]). Furthermore, it surpassed the SOTA Gemini 3.1 Pro model (Accuracy 90% [95% CI: 83-95%]; Sensitivity 86% [95% CI: 74-95%]). While the HITL sensitivity demonstrated a definitive and statistically significant edge over the Gemini model, the accuracy improvement fell just slightly short of undisputed statistical significance due to overlapping confidence intervals. Conclusions: By utilizing their clinical domain knowledge of tumor invasion patterns and topological priors, surgeons effectively filtered algorithmic noise-overriding ML errors in 69% (9 out of 13) false positive cases that models alone could not resolve. This demonstrates exactly how and where HITL optimally utilizes human contextual intelligence to outperform autonomous "models-only" pipelines, confirming a human-ML synergy that augments the objectivity of machine learning with human domain knowledge. This paradigm ensures that the ultimate responsibility for diagnostic inference remains safely and practically in human hands. Open Data Initiative: To ensure essential reproducibility, enable independent multi-center validation and support open science, all examples of intraoperative in vivo OCT brain scans used in this study are made publicly available. To the best of our knowledge, this represents the first open-access data of its kind globally.