The Establishment And Validation Of The Deep Learning Based Diagnostic System For Cervical Squamous Intraepithelial Lesions In Colposcopy

As the golden standard to diagnose cervical cancer and its precancerous lesions,colposcopy guided biopsy faced many limitations.The accuracy,sensitivity and specificity of colposcopy vary with the experience of colposcopists and the patients’ physical condition.In the meantime,the number of patients waiting for colposcopy examination far exceeds the number of experienced colposcopists,making it a long time to wait for colposcopy examination and the following treatment.Medical artificial intelligence and computer-assisted diagnosis can help detect lesions and improve the diagnostic accuracy by using deep learning and the medical image processing technology plus some possible physiological and pathological knowledge,making a breakthrough in the field of medical radiology.The combination of artificial intelligence and colposcopy diagnosis might efficiently ease the pressure of cervical cancer screening,and improve the diagnostic efficiency and accuracy.The research enrolled totally 22330 qualified cases in Women’s Hospital,School of Medicine,Zhejiang University from Aug 2013 to Feb 2019.For each qualified case,her colposcopy images including one saline image,one acetic image and one iodine image at the magnification of 7.5 were collected,as well as the corresponding clinical data including the patient’s age,results of HPV testing and cytology,type of transformation zone(TZ),and pathologic diagnosis.Using the pathologic results as the golden standard,lesion annotations were made by Labelme software.All the cases were randomly divided into three sets of the training set,the valid set and the test set with the ratio of 8 :1:1.In order to get higher efficiency,a pre-trained deep learning model was obtained by training a ResNet model from a database called Image Net,which contains more than 1 million images of over 1000 categories.On that basis,colposcopy images were input to fine-tune multi-modal ResNet classification model,U-Net segmentation model and Mask R-CNN detection model,which use the pre-trained ResNet model as backbone.The sensitivity,specificity,accuracy,positive prediction value,negative prediction value,area under receiver operating characteristic curve,recall,Dice and other indexes were calculated to evaluate the diagnostic level of the models.Furthermore,to compare the different performance of the models in ordinary colposcopy images and in high definition colposcopy images horizontally,an independent dataset of 5384 cases collected in Women’s Hospital,School of Medicine,Zhejiang University from Mar 2019 to Sep 2019 were used to evaluate the three models as an independent validation dataset.Furthermore,colposcopy diagnosis,biopsy sites with pathologic results of five colposcopists with different years’ experience were also collected to be compared with the diagnostic results of models vertically.Through deep learning,the accuracy,sensitivity and specificity of ResNet classification model to differentiate normal cases with squamous intraepithelial lesion(SIL)cases were 84.10%、85.38% and 82.62% respectively,with an AUC of 0.93.On that basis,the recall and Dice of U-Net segmentation model to segment suspicious lesions apart from the cervix were 84.73% and 61.64% in acetic images,with an total accuracy of 95.59%.In iodine images,the recall,Dice and total accuracy were 87.78%,63.80% and 95.70%,respectively.Besides,the sensitivity of Mask R-CNN model to detect high grade squamous intraepithelial lesions(HSILs)was 84.67% in acetic images and 84.75% in iodine images.In the validation of the independent dataset of high definition colposcopy images,the performance of the classification model was poorer than that in ordinary colposcopy images.The sensitivity,specificity and AUC of the classification model were 73.37%、58.16% and 0.71.Compared with the diagnostic level of colposcopists,the diagnostic results of models in ordinary images were higher than the diagnostic level of five colposcopists,and the results in high definition images were lower than that of the senior colposcopists but higher than that of the junior colposcopists.The performance of segmentation model and detection model were almost the same as that in ordinary images.As for the time needed to make a decision,a colposcopist costs usually minutes to cope with a case while the models cost only seconds from image uploading to diagnosis outputting.If the time of image uploading was not included,the models only needed less than one second to make diagnosis.In conclusion,the research demonstrated that the deep learning based colposcopy diagnostic models could differentiate normal images and SIL images precisely,and segment the suspicious lesions and detect HSIL sites to instruct biopsy taking.In the meanwhile,the computer-assisted diagnostic models can efficiently ease the heavy burden of colposcopy,offer a new strategy for patients triage and biopsy sites recommendation.However,the models performed a bit poorer in high definition images than in ordinary images,which needed promoting and optimizing.Part Ⅰ The Establishment of Deep Learning Based Diagnostic System for Cervical Squamous Intraepithelial Lesions in ColposcopyObjective: To collect large amounts of colposcopy images and the corresponding clinical information.To pre-process the big data in order to make preparation for building deep-learning models.To form a systematic colposcopy diagnostic system with classification model,segmentation model and detection model through deep learning on the basis of big data.Methods: All the cases were collected in Women’s Hospital,School of Medicine,Zhejiang University from Aug 2013 to Feb 2019.The unqualified cases were excluded through the exclusion criteria.For each qualified case,her colposcopy images including one saline image,one acetic image and one iodine image at the magnification of 7.5 were collected,as well as the corresponding clinical data including the patient’s age,results of HPV testing and cytology,type of transformation zone(TZ),and pathologic diagnosis.Using the pathologic results as the golden standard,lesion annotations were made by Labelme software.All the cases were randomly divided into three sets of the training set,the valid set and the test set with the ratio of 8 :1:1,and only the results of the valid set were listed in the research.A pre-trained deep learning model was obtained by training a ResNet model from a database called Image Net.On that basis,colposcopy images were input to fine-tune multi-modal ResNet classification model,U-Net segmentation model and Mask R-CNN detection model,which use the pre-trained ResNet model as backbone.The accuracy,sensitivity,specificity,positive prediction value,negative prediction value of the classification model,the recall and Dice of the segmentation model,the accuracy of the detection model were calculated.Results: 1.Through exclusion criteria,totally 22 330 cases were enrolled into the research,including 10 365 normal cases,6 357 low grade squamous intraepithelial lesion cases and 5 608 high grade squamous intraepithelial lesion cases.2.The age distribution in all cases were as follows: women aged between 25 and 55 consisted 92.22% of all the patients,while those who aged below 25 occupied 1.72% and those aged above 55 made up 6.04% of all cases.The infection rate of high risk HPV in all the cases was 94.49%.The distribution of all the cytology results was: NILM:ASCUS:LSIL:ASCH:HSIL:SCC= 33.67%: 26.41%: 23.49%: 7.93%: 8.40%: 0.10%。The transformation zone distribution was as follows: type 1 TZ 12.99%,type 2 TZ 8.10% and type 3 TZ 78.91%.3.The total accuracy,sensitivity and specificity of the ResNet classification model to judge whether the case contains SIL or not were 84.10%,85.38% and 82.62%.The AUC was 0.93.4.The recall and Dice of the U-Net model to segment suspicious lesions were 84.73% and 61.64,with a total accuracy of 95.59% in acetic images.In iodine images,the recall,Dice and total accuracy were 87.78%,63.80% and 95.70% respectively.5.The sensitivity of the Mask R-CNN model to detect HSIL was 84.67% in acetic images and 84.75% in iodine images.Conclusions: The deep learning based diagnostic system can precisely differentiate normal cases and SIL cases,segment suspicious lesions,and assist biopsy taking for HSIL detection.Part Ⅱ The Validation of Deep Learning Based Diagnostic System for Cervical Squamous Intraepithelial Lesions in Colposcopy in the independent datasetObjective: To validate the classification model,segmentation model and detection model in the independent dataset of high definition colposcopy images.To horizontally compare the performance of models in ordinary colposcopy images with that in high definition colposcopy images.To vertically compare the diagnostic level of models with that of clinical colposcopists.Methods: All cases of high definition colposcopy images were collected in Women’s Hospital,School of Medicine,Zhejiang University from Mar 2019 to Sep 2019.Using the same exclusive criteria as the first part,an independent validation dataset of high definition images were formed.The same clinical information as the first part were collected in every case.Besides,the diagnostic results,biopsy sites and the corresponding pathologic results of 5 colposcopists with different colposcopy experience were collected.The images with squamous intraepithelial lesions were also annotated by Labelme software using pathologic results as the golden standard.All the enrolled cases were used as the valid set.The accuracy,sensitivity,specificity,PPV,NPV,recall and Dice were calculated and compared with that of ordinary images horizontally and with that of the colposcopists vertically.Results: 1.The total accuracy,sensitivity and specificity of the ResNet classification model to judge whether the case contained SIL or not in high definition colposcopy images were 63.83%,73.37% and 58.16%.The AUC was 0.71.2.The recall and Dice of the U-Net model to segment suspicious lesions were 85.35% and 47.21%,with a total accuracy of 94.32% in high definition acetic images.In high definition iodine images,the recall,Dice and accuracy were 85.87%,48.74% and 94.52%.3.The sensitivity of the Mask R-CNN model to detect HSIL in high definition images was 84.76% in acetic images and 82.61% in iodine images.The detection model can detect 90.56% HSIL patients through acetic high definition images and 89.78% HSIL patients through iodine high definition images.4.Using SIL as triage,the average sensitivity,specificity and accuracy of five colposcoists were 70%,72.92% and 71.83%.The average PPV and NPV of five colposcopists were 85.02% and 83.03%.The average diagnostic level of five colposcopists were lower than that of the multi-modal classification model in ordinary images and were higher than that of the classification model in high definition images 5.Among SIL cases,the average accuracy of five colposcopists to detect HSIL through biopsy was 27.5%;the average accuracy to detect SIL was 67.97%;the average biopsy number in every case was 2.39.The accuracy of models to detect HSIL was lower than that of colposcopists.Conclusions: The results of three models in high definition images were a bit poorer than that in the ordinary images.The diagnostic level of the models in the ordinary images was higher than that of five colposcopists.The diagnostic level of models in high definition images reached the same level as the junior colposcopists and was lower than that of the senior colposcopists.