Retrospective study demonstrates high-accuracy extraction of cancer recurrence and metastasis from CT reports using semisupervised deep learning, highlighting effective uncertainty modeling.
Key Points
To develop a semisupervised deep learning framework that captures diagnostic uncertainty by classifying postoperative cancer recurrence and metastasis from CT reports into positive, negative, and uncertain categories.
Retrospective analysis of 288,076 postoperative CT reports from 86,083 cancer surgery patients across 11 cancer types at Asan Medical Center (2014–2021).
Model training and evaluation utilized 17,846 deduplicated reports for recurrence and 63,766 for metastasis, combining rule-based algorithms with medical BERT models (MedEmbed and PubMedBERT).
Implemented a human-in-the-loop validation framework across 3 cycles (<1% expert review, ~2,000 samples per cycle) alongside preprocessing with keyword filtering, unsupervised clustering, and evaluation via Integrated Gradients and maximum mean discrepancy testing.
The overall framework achieved multiclass and binary accuracies of 97.33% and 99.33% for recurrence, and 95.00% and 96.67% for metastasis, aligning with clinician intrarater consistencies of 96.88% and 93.80%.
Under simulated real-world conditions, PubMedBERT attained 92.58% multiclass accuracy for recurrence, whereas MedEmbed attained 93.25% binary accuracy for metastasis.
The model successfully captured diagnostic uncertainty in 1.4% of recurrence cases and 6.9% of metastasis cases, with rule-based algorithms outperforming several deep learning architectures in metastasis classification.