Cross Validation of a Conditional Generative Adversarial Network for the Classification of Pleural Mesothelioma Histological Subtype on CT Scans
Abstract
Purpose
This study aimed to cross validate a conditional generative adversarial network (cGAN) model for noninvasive histological subtype classification of pleural mesothelioma on computed tomography (CT) scans.
Methods
CT scans from 70 mesothelioma patients with epithelioid histology and 7 mesothelioma patients with sarcomatoid histology were retrospectively collected. A single axial series was selected from each scan to ensure consistency in reconstruction filter and slice thickness, and a radiologist indicated a single slice that contained the largest cross section of disease. This slice and its two inferior slices were designated as the “Primary Input”, while the three superior slices were designated as the “Secondary Input”. The 70 epithelioid scans were partitioned into 10 folds of 7 scans each. A cGAN, consisting of a generator and discriminator, was trained on 9 folds with augmentations consisting of 5° counterclockwise and clockwise rotations and mirror transformations. The generator was trained to generate the Secondary Input based on the Primary Input, yielding a “Synthetic Output”. The trained generator was tested on the remaining fold and the sarcomatoid cases by generating Synthetic Outputs. L2 loss between the Synthetic Output and the Secondary Input of the test cases was used as the classification metric. Folds were iteratively swapped until each was tested by its respective generator. Model performance was quantified for each fold via receiver operating characteristic (ROC) analysis, and the average area under the curve (AUC) and 95% confidence interval (95% CI) were calculated across folds.
Results
The model achieved an average AUC value of 0.73 (95% CI: [0.68, 0.79]), with a maximum fold AUC value of 0.88.
Conclusion
The model showed potential for histological subtype classification of pleural mesothelioma on CT scans. Further studies will expand the dataset, apply additional data augmentations, and evaluate classification performance between epithelioid and biphasic cases.