Cancer Cell Recognition And Detection Based On Deep Learning

Squamous cell carcinoma is one of the common tumors with a very high mortality rate.Early detection of squamous cell carcinoma lesions can greatly increase the survival chances of patients.The depth learning method has made certain achievements in the recognition and detection of pathological images of squamous cell carcinoma.The use of depth learning technology can help doctors to quickly and accurately identify tumor cell images and assist in diagnosis.However,algorithms based on depth learning usually require a large amount of data training and need to stack a deeper network,and the scale cell image is different from the general natural image in that it has more impurities,the boundaries of different types of cells are not obvious,and the cells are arranged disorderly and irregularly,making it unable to be recognized directly by the existing network.The image quality and resolution of squamous cells are also different.The work of this paper focuses on the effective recognition of squamous cell hierarchy in high resolution and the accurate division of squamous cell image region in low resolution.For squamous cell high-resolution images,an improved TYOLO model based on YOLOV5 is proposed.Transformer method is introduced to add self-attention mechanism,improve the receptor field,carry out multi-scale fusion,strengthen the feature extraction of squamous cell images.The model is embedded into Mosaic data enhancement technology to improve its robustness and recognition ability.Compared with the original network,the improved method can achieve effective feature extraction in the network without deep stacking.In the self-made pathological image data set of squamous cell cells,the m AP value of the improved TYOLO method reached 77.8%,which was significantly improved compared with the traditional YOLOV5 method.Because it is difficult to obtain accurate internal features of squamous cells from low resolution images,an improved TUNET model based on UNET is proposed for low resolution images,in which the Transformer module is added.Combined with the lightweight features of UNET and the advantages of ski connection to retain the original features,and with the help of the multi head self attention mechanism in Transformer,it can effectively capture long-distance dependencies to increase the performance of UNET.It can directly process the original image instead of the feature map extracted by UNET.It has better segmentation performance for small eyes.The Mi OU index of the improved TUNET model for the self-made data scale cell image set is 85.16%,which is greatly improved compared with the UNET method.