Lung cancer treatment gets on the stage of precision medicine. Histopathological classification of lung cancer is crucial in determining optimum treatment. Artificial intelligence (AI) models have been widely shown to be useful in pathological diagnosis and we previously established a reliable AI model to detect the presence of lung cancer on whole slide images (WSIs). However, AI models for the differentiation of major histological types of lung cancer, such as adenocarcinoma (ADC), squamous cell carcinoma (SCC) and small-cell lung cancer (SCLC), are yet to be established.
We trained a convolution neural network (CNN) based on the EfficientNet-B1 architecture to classify ADC, SCC, SCLC, and non-neoplastic lesion from biopsy specimen WSIs (70, 23, 12 and 171 specimens with ADC, SCC, SCLC and non-neoplastic lesion, respectively) using a training dataset of 276 images of which 60 were reserved for validation. The WSIs were manually annotated by pathologists by drawing around the regions that contain each subtype. We used a transfer learning approach, in which the starting weights were obtained from a pre-trained model on ImageNet. The model was then trained on our dataset using a supervised learning approach. To classify a WSI, the model was applied in a sliding window fashion with an input tile size of 224x224 and a stride of 128 on a magnification of x10. The maximum probability was then used as a WSI diagnosis.
We evaluated our model on a total of 533 WSIs that only had WSI diagnoses. The model achieved a Receiver Operator Curve Area Under the Curves of 0.888 (CI 0.872-0.9075), 0.8913 (CI 0.8596-0.9221), 0.9526 (CI 0.9276-0.9646) for ADC, SCC, and SCLC, respectively.
The obtained results on a large test set are a promising first step towards developing a model for the classification of lung cancer. Our model was only trained on a small dataset of 276 WSIs; however, we hope that the model would be further improved with the collection of additional annotated WSIs for training. Having a high performing model could help reduce the burden on pathologists and be useful for the decision of optimum treatment strategies, such as molecular-targeted therapy, immunotherapy and chemotherapy, according to the histological types of lung cancer.
The authors.
Has not received any funding.
All authors have declared no conflicts of interest.