Classification of Cancer Pathology Reports: A Large-Scale Comparative Study

Area of Science:

• Oncology informatics
• Computational pathology
• Medical data science

Background:

• Accurate coding of cancer reports is crucial for epidemiology and treatment.
• Manual coding is time-consuming and prone to errors.
• Automating the classification of International Classification of Diseases for Oncology, 3rd Edition (ICD-O-3) codes is needed.

Purpose of the Study:

• To apply deep learning for automatic and interpretable ICD-O-3 topography and morphology code assignment.
• To evaluate model performance on a large, real-world dataset of Italian cancer reports.
• To compare different deep learning architectures for accuracy and interpretability.

Main Methods:

• Utilized state-of-the-art deep learning techniques.
• Trained models on over 80,000 labeled and 1.5 million unlabeled Italian cancer reports.
• Compared hierarchical, flat, and attention-based models, including element-wise maximum aggregators.

Main Results:

• Achieved 90.3% multiclass accuracy for topography site assignment.
• Achieved 84.8% multiclass accuracy for morphology type assignment.
• Element-wise maximum aggregator models slightly outperformed attentive models and showed interpretability.

Conclusions:

• Deep learning effectively automates ICD-O-3 code assignment for cancer reports.
• Flat models with maximum aggregators provide accurate and interpretable classification.
• This approach can improve the efficiency and consistency of cancer data management.