Automated breast cancer staging from Chinese-language PET/CT reports using two large language models: a retrospective single-center study

Objective To evaluate the performance of GPT-5.5 Pro and DeepSeek-V4-Pro in automated breast cancer cTNM and AJCC anatomic staging from Chinese-language PET/CT reports. Methods A curated test set of 298 pretreatment 18 F-FDG PET/CT reports from routine clinical practice, obtained between February and December 2025 at our institution, was included. All reports had been verified against the corresponding imaging findings and screened to exclude errors or omissions. Following AJCC 8th-edition cTNM criteria, GPT-5.5 Pro and DeepSeek-V4-Pro assigned cT, cN, cM, and AJCC anatomic stage using a task-specific, rule-guided zero-shot prompt developed using a separate set of 20 reports and locked before final evaluation. Model outputs were compared with an expert report-based reference standard. Performance against this reference standard was assessed using accuracy and macro-averaged precision, recall, and F1 score; agreement was assessed using Cohen's κ. Model errors were categorized as diagnostic reasoning, feature processing, or knowledge application errors. Results For cT staging, GPT-5.5 Pro and DeepSeek-V4-Pro achieved accuracies of 98.3% and 97.0% ( P = 0.39) and macro-F1 scores of 0.968 and 0.952 ( P = 0.51), respectively. For cM staging, the corresponding accuracies were 99.7% and 98.0% ( P = 0.06), and the macro-F1 scores were 0.990 and 0.932 ( P = 0.06). cM1 sensitivity was 96.3% versus 77.8%, with 1 versus 6 missed cM1 cases, respectively. A significant performance difference emerged in cN staging, where GPT-5.5 Pro achieved higher accuracy (97.3% vs. 87.2%, P < 0.001) and macro-F1 score (0.959 vs. 0.729, P < 0.001), resulting in superior overall AJCC anatomic stage classification (accuracy, 98.3% vs. 90.3%; macro-F1, 0.968 vs. 0.861; both P < 0.001). Agreement with the reference standard was almost perfect for GPT-5.5 Pro across all tasks ( κ = 0.96–0.98), whereas κ values for DeepSeek-V4-Pro ranged from 0.80 to 0.95. Across the three component TNM tasks, GPT-5.5 Pro produced 14 classification errors, compared with 53 for DeepSeek-V4-Pro; most errors involved cN staging, and DeepSeek-V4-Pro errors were predominantly classified as knowledge application or diagnostic reasoning errors. Conclusion Both models showed high agreement with the expert report-based reference standard under curated input conditions, with GPT-5.5 Pro showing higher performance for cN and overall AJCC anatomic stage classification.