AI ‐based segmentation of occult cervical lymph node metastases: potential reduction of cytokeratin staining

BACKGROUND: Detection of occult cervical lymph node metastases is critical for accurate staging and treatment planning in oral cavity squamous cell carcinoma (OSCC). As part of standard management, elective neck dissection is commonly performed in conjunction with primary tumour resection. While haematoxylin and eosin (H&E) staining is routinely used for histopathological evaluation, small or ambiguous metastatic deposits often require cytokeratin (CK) immunohistochemistry for confirmation. Although CK staining improves sensitivity, it increases diagnostic cost and laboratory workload. Advances in deep learning offer the potential to enhance tumour detection directly from routine H&E sections, particularly in the context of subtle and occult metastatic disease. METHODS: We developed a two-stage deep-learning framework for tumour segmentation using a custom expert-annotated dataset comprising primary OSCC cases and cervical lymph node specimens. The model was pre-trained on primary tumour morphology and subsequently fine-tuned on lymph node metastases to improve detection of small and morphologically subtle tumour deposits. All segmentation masks were generated by a senior pathologist, with CK confirmation incorporated when clinically indicated to guide annotation. A standardized pre-processing pipeline was applied to preserve morphological fidelity while enabling consistent model input. Performance was evaluated on a held-out test cohort using pixel-level segmentation metrics, with additional case-level analysis to assess performance across varying levels of pathological difficulty. Qualitative comparison between AI predictions and CK staining was performed for cases with available immunohistochemistry. RESULTS: During pre-training, the model achieved precision and recall of 0.8484 and 0.8682, respectively. Following fine-tuning on lymph node metastases, performance improved to precision and recall of 0.8778 and 0.8607, respectively. The model demonstrated high overall accuracy (0.9193) and specificity (0.9458), indicating robust discrimination between tumour and non-tumour regions. Case-level analysis revealed significant variability, with lower Dice scores in harder cases, highlighting the increased challenge of accurately delineating occult and micrometastatic disease. In cases with available CK staining, AI-predicted tumour regions exhibited substantial spatial correspondence with CK-positive areas, supporting the model's ability to localize small metastatic deposits using H&E alone. CONCLUSION: AI-based segmentation of H&E sections can accurately localize metastatic deposits in cervical lymph nodes and demonstrate detection of subtle and occult tumour regions. By approximating expert annotations informed by CK, such systems may reduce reliance on immunohistochemistry in selected cases and support more efficient, scalable histopathological workflows.