Salivary gland carcinomas (SGC) are a heterogeneous group of tumors. The prognosis varies strongly according to its type, and even the distinction between benign and malign tumor is challenging. Adenoid cystic carcinoma (AdCy) is one subgroup of SGCs that is prone to late metastasis. This makes accurate tumor subtyping an important task. Matrix-assisted laser desorption/ionization (MALDI) imaging is a label-free technique capable of providing spatially resolved information about the abundance of biomolecules according to their mass-to-charge ratio. We analyzed tissue micro arrays (TMAs) of 25 patients (including six different SGC subtypes and a healthy control group of six patients) with high mass resolution MALDI imaging using a 12-Tesla magnetic resonance mass spectrometer. The high mass resolution allowed us to accurately detect single masses, with strong contributions to each class prediction. To address the added complexity created by the high mass resolution and multiple classes, we propose a deep-learning model. We showed that our deep-learning model provides a per-class classification accuracy of greater than 80% with little preprocessing. Based on this classification, we employed methods of explainable artificial intelligence (AI) to gain further insights into the spectrometric features of AdCys.

Background: Accurate and rapid intraoperative tumor margin assessment remains a major challenge in surgical oncology. Current gold-standard methods, such as frozen section histology, are time-consuming, operator-dependent, and prone to misclassification, which limits their clinical utility. Objective: To develop and evaluate a novel hyperspectral imaging (HSI) workflow that integrates deep learning with three-dimensional (3D) tumor modeling for real-time, label-free tumor margin delineation in head and neck squamous cell carcinoma (HNSCC). Methods: Freshly resected HNSCC samples were snap-frozen and imaged ex vivo from multiple perspectives using a standardized HSI protocol, resulting in a 3D model derived from HSI. Each sample was serially sectioned, stained, and annotated by pathologists to create high-resolution 3D histological reconstructions. The volumetric histological models were co-registered with the HSI data (n = 712 Datacubes), enabling voxel-wise projection of tumor segmentation maps from the HSI-derived 3D model onto the corresponding histological ground truth. Three deep learning models were trained and validated on these datasets to differentiate tumor from non-tumor regions with high spatial precision. Results: This work demonstrates strong potential for the proposed HSI system, with an overall classification accuracy of 0.98 and a tumor sensitivity of 0.93, underscoring the system’s ability to reliably detect tumor regions and showing high concordance with histopathological findings. Conclusion: The integration of HSI with deep learning and 3D tumor modeling offers a promising approach for precise, real-time intraoperative tumor margin assessment in HNSCC. This novel workflow has the potential to improve surgical precision and patient outcomes by providing rapid, label-free tissue differentiation.