Pathological classification of oral cancer based on multi-instance network and two-level attention(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2024年第8期
- Page:
- 946-952
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Pathological classification of oral cancer based on multi-instance network and two-level attention
- Author(s):
- JIANG Huimin; FANG Liming; TAO Long
- Department of Medical Imaging, Wannan Medical College, Wuhu 241000, China
- Keywords:
- Keywords: oral cancer multi-instance learning two-level attention pathological image
- PACS:
- R780.2；R318
- DOI:
- DOI:10.3969/j.issn.1005-202X.2024.08.004
- Abstract:
- Abstract: To address the problems of the low accuracy of pathological classification caused by the large size of pathological data and the high cost of labeling, a pathological classification algorithm for oral cancer is designed based on multi-instance network and two-level attention module, which takes losses at the instance level and image block level into account. A retrospective analysis is conducted on 186 cases of oral cancer (126 cases of squamous cell carcinoma and 60 cases of adenocarcinoma) in the Department of Oral and Maxillofacial Surgery, the First Affiliated Hospital of Wannan Medical College, and the digital pathological sections are divided into training set, verification set and test set. The foreground and background segmentations are performed on the pathological image, and the noise is removed from the background. ResNet50 is used to extract features from the segmented pathological images, and the features are input into the first-level attention network to obtain the attention score and loss based on image block. Then, the image blocks are sorted according to the attention score, and the reset labels are input into the second-level attention network to obtain the loss based on the instance level. The loss of the two-level attention is taken as the total loss of the model, and the prediction result is obtained by training the final network. The experimental results show that the multi-instance network using two-level attention achieves an accuracy of 78.95% and AUC of 0.843 0, demonstrating superior performance than the baseline models.
Last Update: 2024-08-31