Reproduce “Classification of cancer pathology reports: a large-scale comparative study” on OpenResearch

About this paper

Stefano Martina, Leonardo Ventura, Paolo Frasconi

Deep learning models were developed and evaluated for automatically assigning ICD-O3 codes to Italian cancer pathology reports, achieving multiclass accuracies up to 90.3% for topography and 84.8% for morphology. The work also introduced an interpretable model variant, which aids human review by highlighting the specific textual evidence supporting a classification decision.

Key insights

• Flat aggregation methods, specifically element-wise maximum (MAX) pooling, often achieved comparable or superior performance to more complex attention or hierarchical models for classifying these short, keyword-rich pathology reports.
• Hierarchical models did not consistently outperform their flat counterparts, suggesting that the optimal architectural choice depends on the specific textual characteristics and structural complexity of the medical documents.
• Domain-specific word embeddings trained on a large unlabeled corpus of Italian medical text are crucial for achieving high classification performance and capturing the nuances of the specialized sublanguage.

Summary by alphaXiv.