In a single-institution study of 1,668 digitized frozen-section slides from 1,458 surgical patients, an attention-based deep learning model differentiated primary lung cancer from metastatic tumors with an AUC of 0.888 and 80.1% accuracy overall—but squamous cell carcinoma accuracy was just 62.0% intraoperatively. In the study, lung adenocarcinoma performed better at 87.5% accuracy, while colon cancer metastases reached 89.3%, underscoring subtype-dependent performance that could affect real-time surgical decisions.

Why It Matters To Your Practice

• Intraoperative distinction between primary Lung Cancer and metastatic disease can change the operation performed.

• AI support on frozen sections may help when expert pathology resources or time are limited.

• Performance was not uniform across histologies, so clinicians should expect stronger support for some subtypes than others.

Clinical Implications

• Overall performance was reasonable, but squamous cell carcinoma results were notably weaker than adenocarcinoma.

• High accuracy for lepidic (95.5%) and papillary (88.7%) adenocarcinoma patterns suggests some morphologies may be more AI-readable intraoperatively.

• For adenocarcinoma-only comparisons between primary lung cancer and metastases, accuracy was 85.9%; for squamous cell carcinoma, 63.6%.

Insights

• The model used attention-based multiple-instance learning on digitized frozen sections, a workflow aligned with real-world pathology digitization.

• The dataset included 1,170 lung cancer slides and 498 metastatic tumor slides from one institution.

• Further multi-institutional validation is needed before broad adoption, especially for lower-performing subtypes.

The Bottom Line

• AI for intraoperative frozen-section support in Lung Cancer looks promising, but current performance is uneven.

• Clinicians should view these tools as decision support—not a replacement for pathology judgment—particularly when squamous histology is in play.

• The near-term opportunity is selective use in workflows where subtype-specific strengths can improve speed and confidence.