This study aimed to develop a deep-learning model to automatically classify pulmonary nodules based on white-light images and to evaluate the model performance. Besides, suitable operation could be chosen with the help of this model, which could shorten the time of surgery.

All white-light photographs of pulmonary nodules from phones of pathologically confirmed adenocarcinoma in situ (AIS), minimally invasive adenocarcinoma (MIA) and invasive adenocarcinoma (IAC) were retrospectively collected from consecutive patients who underwent surgery between June 30, 2020 and September 15, 2021 at Guangdong Provincial People's Hospital.Finally, a total of 1037 white-light images from 973 individuals were included in the study. The entire dataset was divided into training and test datasets, which were mutually exclusive, using random sampling. Of these, 830 images were used as the training dataset and 104 images from were used as the test dataset. The CNN model was used in classifying images, namely, Resnet-50. For the CNN model, pretrained model with the ImageNet Dataset were adopted using transfer learning. After constructing the CNN models using the training dataset, the performance of the models was evaluated using the test dataset and the prospective validation dataset.

Inclusion Criteria:

1. Male or female，18 years and older.
2. Patients haven't undergone any therapy.
3. The pulmonary nodules were confirmed AIS, MIA or IAC.
4. The sizes of pulmonary nodules were less than 3cm.
5. The images were jpg format.

Exclusion Criteria:

1. Suffering from other tumor disease before or at the same time.
2. Images with poor quality or low resolution that precluded proper classification.

Patients in Guangdong Provincial People's hospital from June 30, 2020 to September 15, 2021.