Effect of Data Augmentation on Classification Performance of Pleural Mesothelioma Histological Subtype on CT Scans
Abstract
Purpose
Accurate histological subtype classification of pleural mesothelioma (PM) from CT imaging remains challenging due to limited labeled data and substantial morphological variability. Data augmentation is a common strategy used to avoid overfitting when training deep learning (DL) models. This study evaluates the effect of individual image augmentation strategies on classification performance and generator reconstruction consistency in a conditional generative adversarial network (cGAN)-based PM subtype classification pipeline.
Methods
A cGAN was trained on 21 CT scans of patients with epithelioid histology. During model training, the following augmentations were applied independently to the data: rotations (±5°, ±10°), Gaussian noise addition, and brightness adjustments. For each augmentation, the model was evaluated using Receiver Operating Characteristic (ROC) analysis on 7 epithelioid and 7 sarcomatoid CT scans unseen by the network. Specifically, an AUC-per-epoch framework logged maximum observed AUC, epoch-wise AUC trajectories, and minimum p-values during training. Generator weights were evaluated post-training to evaluate reconstruction stability per augmentation.
Results
Moderate geometric and intensity-based augmentations produced small but consistent improvements in classification performance relative to the baseline, which achieved a maximum AUC of 0.796. Rotational augmentations showed the greatest improvement, with rotations of −10° achieving an AUC of 0.857. Brightening augmentations brought an AUC of 0.796, while decreases in brightness increased AUC to 0.816, and Gaussian noise was the only augmentation to result in reduced performance (AUC 0.776). Across augmentations, performance gains emerged gradually across epochs, indicating improved robustness rather than early overfitting.
Conclusion
Targeted, physically plausible augmentations modestly improve mesothelioma subtype classification performance in a cGAN-based framework, while noise-driven transformations reduce generalizability. These findings suggest the importance of focusing on the selection of specific augmentation strategies rather than just quantity, as well as that controlled geometric transformations appear to be more beneficial than stochastic perturbations for CT-based tumor classification.