Semisupervised BERT framework classifies cancer recurrence and metastasis from CT reports, matching clinician accuracy

A semisupervised deep learning framework classified postop recurrence and metastasis from CT reports with up to 99.33% accuracy (binary), matching clinician consistency, while explicitly modeling diagnostic uncertainty (1.4% recurrence, 6.9% metastasis).

• Retrospective study of 288,076 postoperative CT reports from 86,083 cancer surgery patients at Asan Medical Center (2014-2021).
• PubMedBERT achieved 92.58% multiclass accuracy for recurrence; MedEmbed 93.25% binary accuracy for metastasis in simulated conditions.
• Framework required expert annotation on <1% of reports; generalizability needs multicenter validation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.