Pre-coding skin cancer from free-text pathology reports using noise-robust neural networks

Area of Science:

• Oncology
• Medical Informatics
• Artificial Intelligence

Background:

• Cancer registries rely on manual coding of free-text pathology reports.
• Skin cancer incidence is rising, posing a challenge for manual coding efforts.
• Accurate coding is crucial for epidemiological studies and cancer surveillance.

Purpose of the Study:

• To develop an AI-based method for automated identification and coding of skin cancer from pathology reports.
• To extract key diagnostic terms and suggest clinical variables for improved coding efficiency.
• To enhance the accuracy and speed of cancer case identification in population-based registries.

Main Methods:

• Exploration of convolutional neural networks (CNNs) with noise-robust loss functions for cancer identification and pre-coding.
• Application of an attention mechanism to highlight key diagnostic terms within reports.
• Training and evaluation using previously registered and manually coded skin cancer cases.

Main Results:

• High accuracy (0.98-0.99) in detecting skin cancer types.
• Accurate pre-coding for subsite, morphology, tumor behavior, and laterality (F1 scores 0.89-0.99).
• Effective extraction of key terms matching ICD-O code descriptions with high precision.

Conclusions:

• The AI method successfully identified and pre-coded skin cancer cases in a pilot study at the Vaud Cancer Registry.
• Medical coders found the AI pre-coding useful, time-saving, and facilitating review.
• Future integration into registry workflows and expansion to other cancer types are planned.