Neural Network Assisted Pathology Case Identification.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Natural language processing (NLP)

Background:

• Traditional cohort selection relies on structured query language (SQL) keyword searches.
• Neural network-based NLP pipelines offer a more accurate alternative for case retrieval.
• Automated case identification is crucial for research and quality assurance.

Purpose of the Study:

• To evaluate the accuracy of a convolutional neural network (CNN) model in identifying primary colon adenocarcinoma from pathology reports.
• To compare the performance of NLP-based methods with traditional SQL searches for pathology case retrieval.

Main Methods:

• 1000 pathology reports containing "colon" and "carcinoma" were retrieved via SQL.
• Reports were manually labeled as primary colon adenocarcinoma (positive) or other (negative).
• A CNN model was trained on a subset of reports and validated on a holdout set.

Main Results:

• The CNN model achieved 92% accuracy in classifying primary colon adenocarcinoma cases.
• The model demonstrated a high area under the ROC curve (AUC) of 0.957.
• Accurate classification was achieved for both positive and negative cases.

Conclusions:

• CNN models can accurately identify specific pathology cases, such as primary colon cancer.
• AI-powered NLP methods can serve as a valuable adjunct or alternative to traditional text extraction techniques.
• This approach improves the efficiency and accuracy of pathology case retrieval for research and quality assurance.