You Position: Home > Paper

Interpretable classification of colon histological images using attention-based residual network model

( views:0, downloads:0 )
No author available
Journal Title:
Chinese Journal of Biomedical Engineering
Key Word:
注意力机制;注意力热图;残差网络;病理学图像;可解释性;Attention mechanism;Attention heatmap;Residual network;Pathological image;Interpretability

Abstract´╝Ü Objective:To establish a residual network model based on attention mechanism for multi-category diagnosis of eight commonly-seen "image patches" in colorectal pathology.Methods:Attention mechanism based residual neural network learning was used for label prediction of eight colon histological images. Meanwhile, the characteristics of interpretable classifications were obtained by decoding the network-learned patterns based on attention heatmaps generated by gradient-weighted class activation mapping (Grad-CAM) .Results:The proposed method achieved a 98.78% accuracy and an AUC area of 0.998 3 in classification of 8 colon histological image patches including colon cancer. The attention heatmaps generated generally coincided with classifications by professional pathologists.Conclusion:The residual network model based on the attention mechanism can achieve high accuracy in the multi-classification of pathological images. This helps physicians in determining on symptomatic treatment in response to real-time, objective and accurate diagnostic findings, and also offers interpretability for classification of histological images.

  • This article has no references!
WanfangData CO.,Ltd All Rights Reserved
About WanfangData | Contact US
Healthcare Department, Fuxing Road NO.15, Haidian District Beijing, 100038 P.R.China
Tel:+86-010-58882616 Fax:+86-010-58882615