Home > News Center > Medical Sciences > The study of deep learning approach to classify lung cancer and its mimics using WSI by Professor Li Weizhong's team was published in BMC Medicine

The study of deep learning approach to classify lung cancer and its mimics using WSI by Professor Li Weizhong's team was published in BMC Medicine

Last updated :2021-04-30

Source: Zhongshan School of Medicine
Edited by: Tan Rongyu, Wang Dongmei

Professor Li Weizhong’s team at Zhongshan School of Medicine and Professor Ke Zunfu’s team at The First Affiliated Hospital of Sun Yat-sen University jointly developed an intelligent diagnostic model for lung histopathology using deep learning technology. The model can accurately distinguish lung cancer and its easily confusing diseases from histopathological images. The study “Deep learning-six-type classifier for lung cancer and mimics from histopathological whole slide images: a retrospective study” was published on March 29, 2021 in BMC Medicine.

The researchers constructed the deep learning classifier of six-type lung diseases from histopathological images through supervised learning, visualized results into heat maps, further validated the model performance using independent data sets from multiple medical centers, and finally evaluated the clinical significance of the model through a human-machine comparison. The model is the first multi-classifiers to distinguish between lung adenocarcinoma (LUAD), lung squamous cell carcinoma (LUSC), non-small cell lung carcinoma (SCLC), organizing pneumonia (OP), pulmonary tuberculosis (PTB), and normal lung (NL), expanding the scope of artificial intelligence-assisted diagnosis to meet more complex diagnostic needs. The researchers tested more than 1000 pathological slices from four different medical centers, with the outcome maximum AUC of 0.978, which was highly consistent with the ground truth of clinical diagnosis. The researchers also invited four pathologists with different clinical experience to conduct a double-blind review on the images, and found the model highly consistent to the experienced pathologists.
With the broad coverage of lung diseases, the rigorous validations on multi-center cohorts, the improved interpretability of the results, and the comparable consistency with experienced pathologists, the model exhibited excellent accuracy, robustness, efficiency, and practicability as a promising assistant tool.



 

 
Figure legend: (a) Visualization heatmaps of tissue predictions of LUAD, LUSC, SCLC, PTB, OP, and NL from left to the right, respectively. (b) Sankey diagram illustrates the difference among ground truth, best pathologist and our six-type classifier.

Paper links:
https://rdcu.be/chEIH
https://bmcmedicine.biomedcentral.com/articles/10.1186/s12916-021-01953-2