How We Trained AI Models to Detect Tumors and Gene Mutations

Researchers trained AI models on TCGA datasets (BRCA & LUSC) for tumor and TP53 mutation detection. Results show varying accuracy across slide types and magnification levels. FFPE slides improved mutation detection, while frozen slides sufficed for tumor detection. Model generalization was tested using cross-validation and optimal sampling techniques.

Table of Links

Abstract and I. Introduction

3. Results

3.1. Training Methods

We split each dataset into a training set (80%) and test set (20%). For each task and model, we performed 5-fold cross-validation on the training set in order to prevent overfitting. The test set was then used for external validation, with the model extracted from the fold that obtained the best results.

For all the models, the loss function used was the binary cross-entropy loss, defined as

where 𝑌 is the positive class. For optimization of the loss, we used the Adam optimiser. For ...