The study examined discrepancies between ground truth TNM classifications from the Kyoto Record and the Kyoto Registry Dataset across various cancers. For cervical cancer, 83% matched in pT classification, with most errors in T1b sub-classifications. Uterine corpus cancer revealed a 94.5% match, while ovarian cancer showed a 86.3% match, with common misclassifications noted. Internal inconsistencies were identified, particularly in cervical cancer’s radiological findings versus clinical classifications. External registry data showed a 3.4% error rate, with variations across institutions. The impact of classification guideline revisions highlighted differing inconsistencies post-revision. Using the cloud-based LLM, Gemini, an accuracy exceeding 99% for pT and pN classifications was achieved, and local models like Qwen2.5 yielded competitive results. Enhanced classification accuracy was facilitated by Pydantic-constrained decoding, proving superior compared to conventional prompting methods. This indicates a need for improved data accuracy in cancer registries, leveraging advanced machine learning techniques.
Source link
Share
Read more