A gastric cancer recognition algorithm on gastric pathological sections based on multistage attention‐DenseNet

Abstract

Summary As an important method to diagnose gastric cancer, gastric pathological sections images (GPSI) are hard and time‐consuming to be recognized even by an experienced doctor. An efficient method was designed to detect gastric cancer in magnified (20×) GPSI using deep learning technology. A novel DenseNet architecture was applied, modified with a multistage attention module (MSA‐DenseNet). To develop this model focusing on gastric features, a two‐stage‐input attention module was adopted to select more semantic information of cancer. Moreover, the pretraining process was divided into two steps to improve the effect of the attention mechanism. After training, our method achieved a state‐of‐the‐art performance yielding 0.9947 F1 score and 0.9976 ROC AUC on a test dataset. In line with our expectation in clinical practice, a high recall (0.9929) was produced with high sensitivity to the positive samples. These results indicate that this new model performs better than current artificial detection approaches and its effectiveness is therefore validated in cancer pathological diagnoses.

Dense Convolutional Network and Its Application in Medical Image Analysis

“…First, multilevel attention mechanism is introduced in transition layer to fuse adjacent dense block output features and generate channel spatial attention map to enhance transition layer features. In diagnosing gastric cancer in gastric pathology section images, although DenseNet rich semantic information can detect larger sensory field data, it is still difficult to obtain spatial information and identify hidden gastric features in feature maps, Liu et al [ 81 ] proposed multilevel attention dense network (MSA-DenseNet) to detect gastric cancer in 20x magnified section images; transition layer introduced attention providing attention vectors to enhance gastric features, selecting more semantic information about cancer, global average pool, and fully connected layer to provide nonlinearity for channel attention, using adjacent dense block features to enhance relevant semantic information, combining lower-order features with more spatial prediction and higher-order features with more semantic prediction features; final model obtains better detection results than existing manual detection methods.…”