Deep Learning Model-Based Architectures for Lung Tumor Mutation Profiling: A Systematic Review.

Area of Science:

• Computational pathology
• Digital histopathology
• Artificial intelligence in oncology

Background:

• Non-small cell lung cancer (NSCLC) presents significant heterogeneity, complicating molecular profiling and treatment.
• Deep learning (DL) offers a novel approach to extract genomic insights from routine Hematoxylin and Eosin (H&E) histopathology images.
• This complements traditional Next-Generation Sequencing (NGS) for precision oncology.

Purpose of the Study:

• To systematically review the application of DL models in predicting molecular alterations in NSCLC using H&E-stained histopathology slides.
• To evaluate the performance and limitations of current DL approaches in this domain.

Main Methods:

• A systematic literature search adhering to PRISMA 2020 guidelines was performed across major scientific databases (PubMed, Scopus, Web of Science).
• Studies published up to March 2025 utilizing DL for NSCLC mutation prediction from H&E slides were included.
• Data extraction focused on model architectures, datasets, and performance metrics.

Main Results:

• Sixteen studies met the inclusion criteria, predominantly using convolutional neural networks (CNNs).
• Models were often trained on The Cancer Genome Atlas (TCGA) dataset to predict mutations like EGFR, KRAS, and TP53.
• Reported predictive performance (Area Under the Curve) varied from 0.65 to 0.95, indicating promising but inconsistent results.

Conclusions:

• DL-based histopathology demonstrates significant potential for correlating tissue morphology with molecular signatures in NSCLC.
• Methodological inconsistencies, limited sample sizes, and insufficient external validation hinder clinical translation.
• Standardization, larger multicenter studies, and robust validation are crucial for future clinical implementation.