A novel transformer-based aggregation model for predicting gene mutations in lung adenocarcinoma

Area of Science:

• Computational pathology
• Genomics
• Artificial intelligence in medicine

Background:

• Predicting gene mutations from whole slide imaging (WSI) is crucial for lung adenocarcinoma research.
• Challenges include extracting global information and unbiased semantic aggregation from complex WSI data.

Purpose of the Study:

• To develop and validate a novel Transformer-based aggregation model for predicting gene mutations in lung adenocarcinoma.
• To mitigate semantic bias and address data limitations in WSI analysis.

Main Methods:

• Proposed a Transformer-based aggregation model with a self-learning weight aggregation mechanism.
• Employed a random patch training method to enhance model learning richness.
• Utilized lung adenocarcinoma datasets from Shandong Provincial Hospital and The Cancer Genome Atlas (TCGA).

Main Results:

• The model demonstrated improved performance in predicting TP53, CSMD3, LRP1B, and TTN gene mutations.
• Achieved an average increase of 4% in the Area Under the ROC Curve (AUC) value.
• Successfully correlated pathological image features with molecular characteristics.

Conclusions:

• The developed model offers an efficient approach for exploring the link between pathological features and molecular characteristics in lung adenocarcinoma.
• Presents a novel method for clinical genetic testing, enhancing the identification of molecular features and genetic testing efficiency.
• Aims to provide more accurate and reliable results for lung adenocarcinoma studies.