JMIR Publications

ABSTRACT

Background:

Despite the success of deep learning (DL) for the automated and early diagnosis of lung cancer, the limited availability of clinical data for training these algorithms restricts their performance. While data augmentation (DA) offers a potential solution, existing DA methods have limitations when applied to chest computed tomography (CT) scans across various DL architectures.

Objective:

This study proposes the Random Pixel Swap (RPS) data augmentation technique to improve the diagnostic capabilities of DL models across convolutional and transformer architectures when applied to lung cancer diagnosis from patient CT scans.

Methods:

The proposed technique generates new data by randomly swapping pixels within a patient’s CT scan. It was validated using pre-activated, state-of-the-art convolutional neural networks (CNNs) and transformer DL models on two publicly available chest CT datasets. The models were evaluated based on accuracy and the area under the receiver operating characteristic curve (AUROC).

Results:

The RPS improved the diagnostic capability of the selected models, outperforming state-of-the-art DA techniques. It achieved an overall best accuracy of 94.51% and 95.77% AUROC on the IQ-OTH/NCCD dataset, 97.78% accuracy, and 99.46% AUROC on the chest CT-scan dataset. These results surpassed those of previous studies that have analyzed these datasets.

Conclusions:

The RPS technique can effectively enhance the performance and diagnostic capability of both CNN and Transformer DL architectures when applied to automate lung cancer diagnosis using patient CT scans.