Award winners

Manual coding of topography and morphology, together with the extraction of clinicopathological variables from histopathological reports (HR), is a time-consuming process in cancer registration. Machine learning (ML) can support this workflow by automating text classification and structured information extraction from routine diagnostic reports. The aims are to develop and evaluate automated models for assigning predefined categories (topography/morphology) to HR using TF-IDF (Term Frequency-Inverse Document Frequency) representations and ML algorithms and to extract clinicopathological variables, including TNM stage, Gleason score, differentiation grade, PGR, Ki67, ER, HER2, Breslow thickness and Clark level, using rule-based and transformer-based approaches. Methods: HR relating to non-multiple incident cases from the Milan Cancer Registry (2019–2021) were included, restricting reports to a time window of -3/+9 months from the incidence date. After text preprocessing and data cleaning, the dataset was split equally for training and validation. Classification models were developed using TF-IDF features and One-vs-Rest logistic regression. Additional variables were extracted primarily through regular expressions and linguistic patterns. For Breslow thickness and Clark level, dedicated transformer-based named entity recognition modules were introduced, fine-tuned on approximately 1,000 annotated examples per variable. Clark level extraction was implemented with a hybrid strategy combining high-precision textual patterns with transformer support for less standardised formulations. Results: The dataset comprised 45,091 HR. On the test set, the model achieved 95% accuracy for topography and 83% for 4-digit morphology. For breast, lung and prostate tumours, accuracy approached 100%. Extraction of additional clinicopathological variables showed concordance rates above 95%, while Breslow and Clark modules extended the system’s ability to capture clinically relevant melanoma-specific information from free text. Conclusions: ML applied to HR can automate l coding and support the extraction of clinicopathological variables. The integration of traditional text classification models with rule-based and transformer-based extractors can provide effective support for Cancer Registry activities.