| Title |
Development of Lung Cancer Computer-aided Diagnosis System using Convolution Neural Network based on Deep Learning Model |
| Authors |
박재범(Jae-beom Park) ; 신소영(So-young Shin) ; 조현종(Hyun-chong Cho) |
| DOI |
https://doi.org/10.5370/KIEE.2023.72.11.1406 |
| Keywords |
CADx; Cancer Diagnosis; Classification; Convolution Neural Network; Deep learning; Lung cancer; Malignant Tumor |
| Abstract |
Lung cancer ranked second in Korea domestic cancer incidence in 2020 and second in death rate. Lung cancer often has no early symptoms, so patients often miss the time of treatment. Accordingly, in Korea, lung cancer has been included in the national cancer screening since 2019. However, among misdiagnosis cases, lung cancer had the highest misdiagnosis rate, and the accuracy of screening may vary depending on the medical specialist's skill level and fatigue. Accordingly, this paper proposed a lung cancer CADx(Computer-Aided Diagnosis) system based on EfficientNetV2-L and ConvNeXt-B. EfficientNetV2 is a model that can have high classification performance with a small number of parameters using the Training-Aware NAS (Neural Architecture Search) method. ConvNeXt is a network that achieves higher performance than ViT(Vision Transformer) by combining the latest techniques with ResNet-50 as a base model. Medical imaging generally suffers from a data shortage problem. Therefore, we augmented the lung cancer dataset using AutoAugment using the ImageNet augmentation policy. Through this method, the sensitivity in classifying malignant(lung cancer) and normal improved from 0.8354 to 0.9638 in EfficientNetV2 and from 0.9796 to 0.9963 in ConvNeXt.
AUC (Area Under the ROC Curve) also improved from 0.9967 to 0.9974 for EfficientNetV2 and from 0.9973 to 1.0000 for ConvNeXt. Additionally, noise that may generally occur in CT images was added and compared through Gaussian noise.
EfficientNetV2's Sensitivity was 0.7417 in the original model and 0.8954 in the model to which AutoAugment was applied, representing a decrease of 9.37% and 6.84%, respectively. In contrast, ConvNeXt exhibited a Sensitivity of 0.9796 in the original model and 0.9963 in the model to which AutoAugment was applied, showing no decrease in performance. This led to the development of a CADx system that demonstrates excellent performance. |