Machine learning-assisted monitoring of radiofrequency ablation in hepatic tumors using spectral traces.

PURPOSE: Accurate intraoperative assessment of radiofrequency ablation (RFA) margins in hepatic malignancies remains a significant clinical challenge. METHODS: This study investigates the efficacy of hyperspectral imaging (HSI) for objective, spatially resolved discrimination of ablated tissue, viable tumor, and normal parenchyma in 10 ex vivo liver specimens. Analysis of visible-near-infrared (VNIR) reflectance revealed distinct spectral signatures, characterized by a consistent monotonic hierarchy (ablated < normal < tumor), enabling reliable tissue differentiation. Segmentation performance was systematically evaluated using four supervised classifiers: Spectral Angle Mapper (SAM), Support Vector Machine (SVM), Random Forest (RF), and Linear Discriminant Analysis (LDA). The total delineated tissue area exhibited notable inter-method variability, ranging from 137,283 to 189,289 pixels (+ 37.9%), underscoring sensitivity to classifier architecture. RESULTS: Nevertheless, all approaches maintained biologically coherent tissue distributions, with normal parenchyma predominating (44-56%), followed by tumor (25-39%) and ablated regions (17-20%). That is consistently indicating the necessity for monitoring ablative margins. SAM, SVM, and LDA classifiers achieved class-specific detection errors around 3.5% and 18%. LDA attained the highest overall accuracy (94.5%), demonstrating the potential of HSI for reliable, intraoperative image-guided tissue characterization. Dimensionality reduction across 462 spectral bands identified an optimal discriminative window within 643-698 nm, supporting the development of streamlined multispectral systems. CONCLUSION: Overall, HSI is neither a contact nor destructive modality for quantitative RFA assessment. Despite classifier-dependent variability necessitating standardized or ensemble strategies, these findings support its prospective integration into real-time surgical guidance, pending in vivo validation.