Deep learning on H&E localised occult neck node metastases from oral cancer

Design

Deep-learning model development with held-out test set

Population

Expert-annotated OSCC primary and cervical lymph node slides

Primary outcome

Pixel-level segmentation of metastatic tumour

Effect

Precision 0.88, recall 0.86, specificity 0.95; weaker on occult deposits

Occult nodal disease in oral squamous cell carcinoma changes staging and adjuvant treatment, and small deposits often need cytokeratin IHC to confirm. This group trained a two-stage model, first on primary tumour and then on nodal metastases, with masks drawn by a senior pathologist informed by cytokeratin.

On a held-out test set the model reached precision 0.88, recall 0.86, accuracy 0.92 and specificity 0.95 at pixel level. Performance fell in harder cases, with lower overlap scores for occult and micrometastatic disease. Where cytokeratin was available, predicted regions corresponded well with positive areas.

This is early development work. Pixel-level metrics do not tell you how many node-positive patients would be missed, which is the number that matters for a screening tool.

• Keep cytokeratin IHC for ambiguous nodal deposits; AI is not yet a substitute
• Performance was weakest exactly where help is most needed — micrometastases
• Ask for case-level sensitivity before adopting any nodal screening algorithm
• Annotated in-house datasets are valuable; record cytokeratin-confirmed cases

Why it matters

Cytokeratin staining of every neck dissection is costly, and AI triage could target it.

Don't overread it

Pixel-level metrics from a single development set do not show clinical sensitivity per patient.

Read the rest in the app

You have read your two free briefings this month. The app carries all 27 specialties, every morning, free — and this finding is waiting in it.

Get Daily Doctor — free

Scan to keep reading on your phone. No account needed to start.