Objective:To study the association between histopathological features and HER2 overexpression/amplification in breast cancers using deep learning algorithms.Methods:A total of 345 HE-stained slides of breast cancer from 2012 to 2018 were collected at the China-Japan Friendship Hospital, Beijing, China. All samples had accurate diagnosis results of HER2 which were classified into one of the 4 HER2 expression levels (0, 1+, 2+, 3+). After digitalization, 204 slides were used for weakly supervised model training, and 141 used for model testing. In the training process, the regions of interest were extracted through cancer detected model and then input to the weakly supervised classification model to tune the model parameters. In the testing phase, we compared performance of the single- and double-threshold strategies to assess the role of the double-threshold strategy in clinical practice.Results:Under the single-threshold strategy, the deep learning model had a sensitivity of 81.6% and a specificity of 42.1%, with the AUC of 0.67 [95% confidence intervals (0.560,0.778)]. Using the double-threshold strategy, the model achieved a sensitivity of 96.3% and a specificity of 89.5%.Conclusions:Using HE-stained histopathological slides alone, the deep learning technology could predict the HER2 status using breast cancer slides, with a satisfactory accuracy. Based on the double-threshold strategy, a large number of samples could be screened with high sensitivity and specificity.

Objective:To establish an artificial intelligence (AI)-assisted diagnostic system for lung cancer via deep transfer learning.Methods:The researchers collected 519 lung pathologic slides from 2016 to 2019, covering various lung tissues, including normal tissues, adenocarcinoma, squamous cell carcinoma and small cell carcinoma, from the Beijing Chest Hospital, the Capital Medical University. The slides were digitized by scanner, and 316 slides were used as training set and 203 as the internal test set. The researchers labeled all the training slides by pathologists and establish a semantic segmentation model based on DeepLab v3 with ResNet-50 to detect lung cancers at the pixel level. To perform transfer learning, the researchers utilized the gastric cancer detection model to initialize the deep neural network parameters. The lung cancer detection convolutional neural network was further trained by fine-tuning of the labeled data. The deep learning model was tested by 203 slides in the internal test set and 1 081 slides obtained from TCIA database, named as the external test set.Results:The model trained with transfer learning showed substantial accuracy advantage against the one trained from scratch for the internal test set [area under curve (AUC) 0.988 vs. 0.971, Kappa 0.852 vs. 0.832]. For the external test set, the transferred model achieved an AUC of 0.968 and Kappa of 0.828, indicating superior generalization ability. By studying the predictions made by the model, the researchers obtained deeper understandings of the deep learning model.Conclusions:The lung cancer histopathological diagnostic system achieves higher accuracy and superior generalization ability. With the development of histopathological AI, the transfer learning can effectively train diagnosis models and shorten the learning period, and improve the model performance.