A BERT model generates diagnostically relevant semantic embeddings from pathology synopses with active learning.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Natural language processing for healthcare

Background:

• Pathology synopses are expert-written summaries of tissue observations crucial for diagnosis.
• Interpreting these synopses is time-consuming and requires specialized knowledge.
• Limited specialist availability hinders the full utilization of information within pathology synopses.

Purpose of the Study:

• To develop a deep learning model for extracting semantic information from pathology synopses.
• To create a set of semantic labels for bone marrow aspirate pathology synopses using active learning.
• To leverage extracted embeddings for improved diagnostic capabilities.

Main Methods:

• An active learning approach was used to define semantic labels for pathology synopses.
• A transformer-based deep learning model was trained to map synopses to semantic labels.
• Learned embeddings were extracted from the model's hidden layer for feature representation.

Main Results:

• Transformer models can extract diagnostically relevant embeddings from pathology synopses with limited training data.
• These embeddings accurately map patients to probable diagnostic groups, achieving a micro-average F1 score of 0.779.
• The model demonstrates the utility of deep learning for information extraction in complex pathology data.

Conclusions:

• A generalizable deep learning model and approach can unlock semantic information in pathology synopses.
• This technology can enhance diagnostics, support biodiscovery, and advance AI-assisted computational pathology.
• The findings pave the way for more efficient and accurate analysis of pathology data.