Automatic Grading Of Penile Squamous Cell Carcinoma Based On Pathological Image Calculations

Squamous cell carcinoma of penis is one of the most common malignant tumors in the male genitourinary system.The malignancy of penile squamous cell carcinoma is usually characterized by “Grading”,which is divided into low,medium and high according to the pathological signs of patients.Grading is crucial to the selection of subsequent treatment options and the assessment of patients' survival,and it is the diagnosis that pathologists must make when reading the slices.Moreover,it is also a regional indicator,that is,a slice may contain multiple grades of regions.Before making the diagnosis,the pathologist need to review all the cancer areas,analyze the nuclei atypia,mitotic,and other characteristics.Such a diagnosis process is time-consuming,labor-intensive and highly subjective.Therefore,it is important and urgent to develop an automatic diagnosis model based on computational pathology technology to assist doctors in quantitative grading diagnosis.The automatic grading of penile squamous cell carcinoma in this paper mainly includes two parts:(1)Based on the pathologists' reading process,novel deep neural network models were constructed to locate various tissues in the slices and segment the nuclei in tissues.(2)According to the clinical diagnostic basis of pathologists in grading,a series of effective and interpretable histomorphological features were extracted and analyzed for the nuclei in cancer tissues.Finally,a model was constructed to realize the automatic grading of penile squamous cell carcinoma.In the first work,this paper first classified 7 tissues in penile squamous cell carcinoma slices with a fully convolutional network based on Efficient-Net.The network can be trained with small image blocks,and the trained model can be directly used to predict large image blocks.On the dataset of more than 400,000 patches produced in this paper,the accuracy of the model reached almost 100%,which could quickly and accurately locate the 7 tissue types in penile squamous cell carcinoma slices.In terms of nuclei segmentation,a segmentation model based on distance transform maps was proposed to solve the problem of nuclei adhesion in pathological images.Multi-task learning was applied to guide the network to pay attention to the inner nucleus and weaken the edge of the nucleus.Combined with post-processing operation,the segmentation result of removing the adhesion was obtained.The average Dice coefficient,AJI value,precision and Hausdorff distance of the model in the testing set are 0.779,0.643,0.739 and 10.37,respectively,which are all better than other models.In the second work,inspired by the atypia and cleavage signs that pathologists paid attention to when reading slices,the paper located the cancer region and its nuclei based on the segmentation results of previous work.A total of 661 histomorphological features at the level of nuclei and tissues were designed and extracted.Then,Pearson correlation test,KruskalWallis H test,minimum redundancy maximum correlation(m RMR)and recursive feature elimination(RFE)were used to select the best feature subset composed of 2 nuclei shape features,2 nuclei texture features and 7 tissue texture features.Finally,the support vector machine classifiers were trained for automatic distinguishing of low,medium and high image blocks.The accuracy of the model in the training set and the independent testing set reached87.46% and 86.88%,respectively,and it could achieve accurate automatic grading of penile squamous cell carcinoma.The establishment of the automatic grading model for penile squamous cell carcinoma can provide doctors with quantitative,objective and repeatable grading diagnostic support,so that doctors can make the most appropriate treatment plan on this basis,and ultimately benefit patients.