Fine-grained pathomorphology recognition of cervical lesions with a dropped multibranch Swin Transformer

利用多分支 Swin Transformer 对宫颈病变进行精细病理形态学识别

阅读:1

Abstract

BACKGROUND: Achieving intelligent detection and grading of lesion cells in ThinPrep cytologic test (TCT) pathological images is challenging but may provide considerable clinical value in improving the accuracy of early cervical cancer screening. Therefore, we sought to design a rapid and accurate method for the fine-grained detection of cervical lesions in massive TCT images. METHODS: We developed a YoGaNet architecture based on the YOLOv5l network and dropped multibranch Swin Transformer (DMBST) module. This architecture extracts multilevel global and local features from images, whereas the DMBST module directs the network to learn different features and enhances the capability of fine-grained feature extraction of small objects. The performance of YoGaNet was evaluated on the large public Comparison Detector dataset, containing 7,410 cervical microscopy images and 11 lesion cell categories, and a dataset of a clinical study, which retrospectively enrolled 12 patients with 1,514 cervical microscopy images. RESULTS: In the experiments, YoGaNet achieved the highest mean average precision calculated at an intersection over a union threshold of 0.50 (mAP50) (Comparison Detector dataset: 68.6%; clinical dataset: 36.8%). Compared with those for the Comparison Detector dataset provided in the Liang's reported and the baseline model (YOLOv5l), the mAP50 and recall of YoGaNet improved by 22.8% and 3.2% and by 6.2% and 3.3%, respectively. Moreover, YoGaNet provided considerable advantages in detecting atypical squamous cells of undetermined significance (ASC-US), atypical squamous cells, cannot exclude high-grade squamous intraepithelial lesions (ASCHs), low-grade squamous intraepithelial lesions (LSILs), high-grade squamous intraepithelial lesions (HSILs), squamous cell carcinoma (SCC), Candida cells, and flora cells. CONCLUSIONS: YoGaNet improved the fine-grained recognition of cervical cells in massive TCT images as compared with the baseline model and models reported in previous studies and may thus aid in improving early cervical cancer diagnosis. The inference code is available online (https://github.com/coding-with-chen/YoGA-Net-DMBST).

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。