MR Image Classification for Brain Tumor Texture Based on Pseudo-Label Learning and Optimized Feature Extraction

Brain tumors are the deadliest and most difficult to treat of all forms of cancer. Preoperative classification of brain tumors is conducive to the development of corresponding treatment plan. Take pituitary tumors as an example. Precisely judging the image data of pituitary tumor texture before surg...

Descripción completa

Detalles Bibliográficos
Autores principales: Xu, Qianqian, Xu, Huachang, Liu, Jie, Zhou, Mingxia, Li, Min, Xu, Jinhui, Zhu, Hong
Formato: Online Artículo Texto
Lenguaje:English
Publicado: Hindawi 2022
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9001102/
https://www.ncbi.nlm.nih.gov/pubmed/35419084
http://dx.doi.org/10.1155/2022/7746991
Descripción
Sumario:Brain tumors are the deadliest and most difficult to treat of all forms of cancer. Preoperative classification of brain tumors is conducive to the development of corresponding treatment plan. Take pituitary tumors as an example. Precisely judging the image data of pituitary tumor texture before surgery can provide a basis for the selection of surgical plan and prognosis. However, the existing methods require manual intervention, and the efficiency and accuracy are not high. In this paper, we proposed an automatic brain tumor texture diagnosis method for uneven sequence image data. First, for the small sample of pituitary tumor MRI image data, the T1 and T2 sequence data are uneven or missing; we used the CycleGAN model to perform data conversion between different domains to obtain a completely sampled MRI spatial sequence. Then, we used texture analysis+pseudo-label learning to label pituitary tumor data of some unknown labels. After that, we used the improved U-Net model based on CBAM to optimize feature extraction for pituitary tumor image data. Finally, we used the CRNN model to classify the degree of pituitary tumor texture based on the advantages of sequence data. The entire process only needs to provide labels for the entire sequence data, and the efficiency is greatly improved, with an accuracy rate of 94.23%.