Research On Deep Learning-based Solid Tumor Segmentation And Classification Models

Cancer is a critical disease that origins from epithelial region for most solid tumors.Pathologist diagnosis,risk factor assessment,and histomorphological analysis all begin with the identification of epithelial area.Automatic epithelial tissue region detection and analysis are conducive to quickly help tissue pathologists locate these areas and assist pathologists in deep evaluation.The classification of tumor grades is closely related to the diagnosis strategy and prognosis of patients.In the traditional pathological diagnosis,pathologists usually use the naked eye to observe and capture these types of information,which has a large workload and a long detection cycle.Furthermore,the results obtained by different experts vary.Automatic tumor-level classification can be used as a second reader to provide obj ective auxiliary information,reduce workload and improve accuracy for pathologists.This study focuses on how to effectively use deep learning models for automatic segmentation of tumor epithelial tissue from oral squamous cell carcinoma and automatic classification of cancer regions from breast cancer and validate the model on multicenter data to evaluate the generalizability and to evaluate the equivalence of features from the model results and the annotations made by pathologists.In the study of the automatic breast cancer classification system,n=98 whole slide images from 98 patients with breast cancer were classified at tumor level（n=34 cases with low tumor grade and n=64 cases with high tumor grade）.The pre-trained ResNet was used as the classification neural network,and the cross-entropy function was used as the loss function.3-fold validation is repeated 5 times to validate our results.The average accuracy（ACC）on the validation set is 97.57±0.33%,the average accuracy on the internal test set is 97.53±0.41%,and the average AUC（Area Under ROC Curve）on the external test set is 75.91 ±7.60%.In the study of automatic epithelial tissue segmentation,including a total of n=690 oral cancer histopathological images,an automatic epithelial region segmentation system based on a U-Net was constructed,and the equivalence of feature extraction from manually marked epithelial regions and automatic segmentation regions was evaluated.In the validation set,the model trained with 10x magnification provides the best segmentation results.Compared with the pathologist’s annotation,the pixel accuracy is 88.05%,the recall rate is 82.74%,the precision is 86.41%,and the Dice coefficient is 84.53%.We locked the model trained on 10x magnification as the final model and used it to test and validate the equivalence of morphologic features in an independent test set.It is proved that the feature extraction from the manually annotated epithelial region and the automatically segmented epithelial region is equivalent,that is,most of the feature differences are not significant（p>0.05）.In this study,according to the study of automatic classification of grades of breast cancer and automatic segmentation of oral cancer epithelial region,the experimental results show that the deep learning algorithm can provide pathologists with automatic and objective quantitative classification and segmentation results and improve the diagnostic efficiency and accuracy.It also shows the possibility of the application of deep learning in tumor diagnosis and can ground a foundation for the construction of computer-aided diagnosis and treatment systems in the future.