Combination of generative adversarial network and convolutional neural network for automatic subcentimeter pulmonary adenocarcinoma classification

Background: The efficient and accurate diagnosis of pulmonary adenocarcinoma before surgery is of considerable significance to clinicians. Although computed tomography (CT) examinations are widely used in practice, it is still challenging and time-consuming for radiologists to distinguish between different types of subcentimeter pulmonary nodules. Although there have been many deep learning algorithms proposed, their performance largely depends on vast amounts of data, which is difficult to collect in the medical imaging area. Therefore, we propose an automatic classification system for subcentimeter pulmonary adenocarcinoma, combining a convolutional neural network (CNN) and a generative adversarial network (GAN) to optimize clinical decision-making and to provide small dataset algorithm design ideas.

Methods: A total of 206 nodules with postoperative pathological labels were analyzed. Among them were 30 adenocarcinomas in situ (AISs), 119 minimally invasive adenocarcinomas (MIAs), and 57 invasive adenocarcinomas (IACs). Our system consisted of two parts, a GAN-based image synthesis, and a CNN classification. First, several popular existing GAN techniques were employed to augment the datasets, and comprehensive experiments were conducted to evaluate the quality of the GAN synthesis. Additionally, our classification system processes were based on two-dimensional (2D) nodule-centered CT patches without the need of manual labeling information.

Results: For GAN-based image synthesis, the visual Turing test showed that even radiologists could not tell the GAN-synthesized from the raw images (accuracy: primary radiologist 56%, senior radiologist 65%). For CNN classification, our progressive growing wGAN improved the performance of CNN most effectively (area under the curve =0.83). The experiments indicated that the proposed GAN augmentation method improved the classification accuracy by 23.5% (from 37.0% to 60.5%) and 7.3% (from 53.2% to 60.5%) in comparison with training methods using raw and common augmented images respectively. The performance of this combined GAN and CNN method (accuracy: 60.5%±2.6%) was comparable to the state-of-the-art methods, and our CNN was also more lightweight.

Conclusions: The experiments revealed that GAN synthesis techniques could effectively alleviate the problem of insufficient data in medical imaging. The proposed GAN plus CNN framework can be generalized for use in building other computer-aided detection (CADx) algorithms and thus assist in diagnosis.