Deep Learning for Differentiating Pulmonary Metastasis from Primary Lung Cancer Constructed on Frozen Sections for...

The surgical procedures for primary lung cancer and metastatic lung tumors differ. An intraoperative diagnosis is important for taking decisions regarding the surgical procedure. This study digitized frozen sections for intraoperative diagnosis and developed and evaluated deep learning models to differentiate lung cancer from metastatic tumors. A total of 1668 slides from 1458 patients with lung cancer or metastatic tumors who underwent surgery with an intraoperative diagnosis at a single institution were included. Lung cancer comprised 1170 slides, whereas metastatic tumors comprised 498 slides. Models were constructed to calculate the prediction probability using attention-based multiple-instance learning. Diagnostic performance was assessed using accuracy and area under the curve for each histologic type; the area under the curve was 0.888 (95% CI, 0.871-0.904) with an accuracy of 80.1%. Lung adenocarcinoma prediction accuracy was favorable at 87.5% (95.5% for the lepidic pattern and 88.7% for the papillary pattern), whereas that for squamous cell carcinoma was 62.0%. Colon cancer was the most common metastatic tumor, with an accuracy of 89.3%, followed by soft tissue sarcoma with an accuracy of 82.6%. When limited to adenocarcinoma of lung cancer and metastatic tumors, the accuracy was 85.9%, whereas that for squamous cell carcinoma was 63.6%. The results indicate that deep learning can provide reasonable support in differentiating metastatic lung tumors from primary lung cancer on frozen sections, with performance varying by histologic subtype. Deep learning may assist in intraoperative decision-making under time and resource constraints, although further multi-institutional validation is needed.