The study implemented 5-fold cross-validation on an 80%-20% split of datasets for model training and external validation. Binary cross-entropy loss was used for all models, optimized through the Adam optimizer. Two TCGA projects were analyzed: TCGA-BRCA for tumor detection using flash-frozen slides and TCGA-LUSC for gene mutation detection with FFPE slides. Three magnification levels (5x, 10x, and 20x) were utilized to explore correlations between mutations and tissue morphology.
The loss function used for all models was binary cross-entropy loss. The Adam optimizer was utilized for optimizing the loss across different tasks.
We constructed datasets from TCGA projects, specifically focusing on TCGA-BRCA and TCGA-LUSC, for tumor detection and gene mutation detection tasks.
For tumor detection, flash-frozen slides were selected despite their limitations, while FFPE slides were preferred for gene mutation detection due to training performance.
Three magnification levels were employed for gene mutation detection – 5x, 10x, and 20x – to analyze correlations between mutation and tissue morphology effectively.
Collection
[
|
...
]