Automated Extraction of Postoperative Cancer Recurrence and Metastasis From Computed Tomography (CT) Reports: Semisupervised Deep Learning Study.

Authors

Abstract

Perioperative computed tomography (CT) imaging is essential for detecting postoperative recurrence and metastasis in cancer patients. However, large-scale automated extraction of oncological outcomes from CT reports remains limited by the unstructured nature of report text and wide variability in reporting styles. Radiology reports frequently contain linguistic ambiguities, including negations, hedging, and expressions conveying diagnostic uncertainty (eg, "cannot exclude recurrence" or "possibly metastatic"). Manual review is labor-intensive and constrains consistent extraction at scale. The inability to systematically account for diagnostic uncertainty represents a major barrier to reliable automated surveillance systems. This study aimed to develop a semisupervised deep learning (DL) classification framework that explicitly captures diagnostic uncertainty by classifying postoperative recurrence and metastasis into 3 categories (positive, negative, and uncertain). This retrospective study analyzed 288,076 postoperative CT reports from 86,083 cancer surgery patients at Asan Medical Center (2014-2021). After exact-match deduplication, model training and evaluation used 17,846 unique reports for recurrence and 63,766 for metastasis. Preprocessing identified presumed negatives through keyword filtering and unsupervised clustering. The semisupervised framework incorporated human-in-the-loop validation across 3 cycles-with clinicians reviewing approximately 2000 samples per cycle (<1% of total reports)-and integrated rule-based algorithms (RAs) and medical BERT (MedEmbed and PubMedBERT). A report-level train-validation split was used, as preprocessing reduces each report to sentence-level fragments that preclude patient-level linkage. Performance was evaluated against RAs and multiple BERT variants under both naive and simulated real-world class distributions. Maximum mean discrepancy testing confirmed distributional integrity of the sampled data. Model interpretability was assessed using Integrated Gradients. The cohort included 11 cancer types, predominantly gastrointestinal (28,516/86,083, 33.1%), hepatobiliary and pancreas (14,715/86,083, 17.1%), and genitourinary (12,959/86,083, 15.1%). Under simulated conditions, PubMedBERT achieved 92.58% accuracy for recurrence in the multiclass, and MedEmbed achieved 93.25% for metastasis in the binary class. The framework achieved accuracies of 97.33% (multiclass) and 99.33% (binary class) for recurrence and 95.00% (multiclass) and 96.67% (binary class) for metastasis, compared with human intrarater consistencies of 96.88% and 93.80% (recurrence and metastasis, respectively, for multiclass), reflecting concordance with the clinician-derived consensus standard. The framework captured diagnostic uncertainty in 1.4% of recurrence cases and 6.9% of metastasis cases. Notably, the RA outperformed several sophisticated DL models in metastasis classification. The proposed framework achieves clinician-concordant classification across the full 288,076-report corpus while requiring minimal expert annotation (<1% of reports). By explicitly modeling diagnostic uncertainty and combining rule-based and DL approaches, it demonstrates the potential for automated cancer surveillance and clinical decision support in real-world settings; however, generalizability to other institutions requires prospective multicenter validation.

Tags

Topics