Study On Differentiation Diagnosis Of Uterine Sarcoma And Degenerated Leiomyoma Based On Clinial And MRI Indicators And Artificail Intelligence

Background:Uterine leiomyomas(UL),were the most common uterine benign mesenchyme oriented tumors.They were found in 20-40% of women of childbearing age and 70-80% of women of perimenopausal age,respectively,and treatment includes surgery and conservative treatment.Uterine sarcoma(US),including mesenchymal origin and mixed epithello-mesenchymal origin,mainly includes uterine sarcoma(u LMS),uterine stromal sarcoma(u ESS),uterine carcinosarcoma(u CS),uterine adenosarcoma(u AS)and other rare species.Although the incidence is low,the prognosis is very poor,and the treatment is mainly surgical treatment.How to differentiate these two groups of diseases is a problem that gynecologist must face to every day.Out of 1000 cases of "UL" operation,3cases of US were found.Treating US as UL will lead to disastrous prognosis.The FDA has reported a case of US with misdiagnosis of UL,and power morcellation in laparoscopic surgery led to enterocelia spread.In recent years,with the widespread clinical application of High intensity focused ultrasound(HIFU)in the treatment of UL,the preoperative identification of the two groups of diseases is of great application value for the case selection of this non-micro-invasive treatment method and the formulation of precise personalized treatment plans.Preoperative differential diagnosis of these two groups of diseases is mostly based on clinical and Magnetic resonance imaging(MRI)indicators,but there is no simple,efficient and practical method to differentiate the two groups,especially when the secondary changes occur to UL and UL becomes degenerated leiomyoma(DL).Previous relevant clinical studies mostly focused on u LMS and UL without secondary changes,while there were few studies on other US types.Although MRI-based studies have gradually focused on the differentiation of US and DL in recent years,they are not closely combined with clinical indicators,and the indicators are complex and inconsistent,which is not conducive for clinical promotion.Therefore,it is particularly important to establish a simple,efficient and easy to popularize diagnostic method based on clinical and MRI related indicators.Objective:1.To explore the differenceses between US,u CS and DL and different subtypes of US.2.To explore the value of clinical and MRI indicator in the differentiation of US,u CS and DL.Establishing statistical model and validating it with external data.3.To evaluate the feasibility and effectiveness of machine learning(ML)model of multiparametric magnetic resonance imaging(mp-MRI)features based transfer learning combining with clinical parameters in differentiation US,u CS and DL and compared with models based on radiomics.Materials and methods:Part one1.Research object169 patients with US and u CS,and 430 patients with DL who were in accordance with the inclusive and exclusive criteria.2.Research methodsThe clinical indicators including: body mass index(BMI),age,onset symptom(abnormal vagina bleeding or discharge,accidental found,or mass effect),status of menopause(pre-or post-menopause),carbohydrate antigen125(CA125),carbohydrate antigen 19-9(CA19-9),human epididymis protein 4(HE4),lactate dehydrogenase(LDH),the white blood cell count(WBC),platelet count(PLT),absolute neutrophil count(NEUT),absolute lymphocyte count(LYM),absolute monocyte count(MONO),ratio of neutrophil-to-lymphocyte(NLR),and systemic inflammation response index(SIRI).3.Statistical methods Statistical analysis was performance with SPSS 26.0.Part two1.Research object114 and 278 patients with histopathology proven US+ and DL who were in accordance with the inclusive and exclusive criteria.32 patients with US+from another center(USE)who were in accordance with the same inclusive and exclusive criteria.2.Research methodsClinical indicators were the same with part one.MRI qualitative indicators including: Signal intensity of solid component of tumors was determined on diffusion weighted imaging(DWI)and delay phase.MRI quantitative indicators including:(1)Volume of tumor;(2)The following four apparent diffusion coefficient(ADC)values were measured by placing region of interest(ROI)on ADC mapping,including: average whole tumor ADC(Avg w ADC),minimum whole tumor ADC(Min w ADC),average solid component ADC(Avg s ADC),and minimum solid component ADC(Min s ADC).3.Statistical methods Statistical analysis was performance with SPSS 26.0.Part three1.Research object86 and 189 patients with histopathology proven US+ and DL were retrospectively reviewed.To ensure that the data were balanced,86 cases were randomly selected from 189 cases of DL group to serve as a corresponding dataset.2.Research methodsNon-image data:(1)Age(year);(2)Menopause status(pre-or postmenopause);(3)Clinical manifestation(1.abnormal vagina bleeding or discharge;2.menstrual change;3.abdominal pain;4.urinary symptom;5.accidental found);(4)Maximum diameter of tumors;(5)ADC values of solid component of tumors.Image data:Images of T2 WI and DWI of all included patients were harmonized using the advanced normalization tool and delineation of ROIs were conducted.Lasso CV was used to select deep-learning or radiomics features,random forest(RF)classifier was used to build the model.All data was divided into training and test set in a 7:3 ratio.In total,five models were constructed:(1)Clinical-ADC(C-A)model: based on 4 non-image features;(2)T2 model: constructed based on 8 T2 WI features;(3)DWI model;constructed based on 8 DWI features;(4)mp-MRI model: constructed based on 16 mp-MRI(T2WI and DWI)features;(5)complex multiparametric(complex mp)model: constructed based on 20 features,including 4 nonimage features and 16 mp-MRI features.3.Statistical methods Statistical analysis was performance with SPSS 26.0.Results:Part one1.Comparison of clinical indicators between u LMS,u ESS and u CS groups demonstrated:(1)Comparisons of age,proportion of onset symptom,menopause status,HE4,WBC,and NEUT demonstrated that no significant difference was found between u LMS and u ESS groups,and significant difference was found between u CS and above two groups(both p values were less than 0.05);(2)Comparison of LDH between three groups demonstrated that no significant difference was found between u LMS and u CS groups,and u ESS group was associated with significant lower value than above two groups(p=0.032);(5)no significant difference was found between three groups in the comparison of NLR and SIRI(p=0.485 and p=0.185,respectively).2.Comparison of clinical indicators between US+ and DL groups demonstrate that,14 out of 15 indicators were with significant differences.Comparison of clinical indicators between US and DL groups demonstrate that,8 out of 14 indicators were with significant differences.3.Univariant analysis was applied to 7 blood indicators with significant difference between US and DL groups.LYM achieved the highest sensitivity,CA125 achieved the highest specificity,HE4 achieved the highest AUC.With the combination of 7 blood indicators,the sensitivity,specificity and AUC were 63.2%,75.5% and 0.72,and only AUC outperforming any individual parameter.4.Comparison of clinical indicators between u LMS,u ESS,u CS and DL groups demonstrated that :(1)In the comparison between u LMS and DL groups,7 out of 15 indicators showed significant difference;(2)In the comparison between u ESS and DL groups 5 out of 15 indicators showed significant difference;(3)In the comparison between u CS and DL groups11 out of 15 indicators showed significant difference.Part two1.In the comparison of clinical and MRI indicators,patients in US+ group were associated with higher age,more commonly in post-menopause women and presented with abnormal vagina bleed or discharge than patients in DL group(all p values were less than 0.001),patients in US+group were associated with significant higher WBC,NEUT,NLR,SIRI and lower LYM than patients in DL group(all p values were less than0.05),solid component of tumors of patients in US+ group were more commonly presented hyperintensity on DWI and hypointensity on delay phase,in the meanwhile,solid component of tumors of patients in DL group were more commonly presented isointensity on both DWI and delay phases(both p values were less than 0.05),patients in US+ groups were associated with lower Avg w ADC,Min w ADC,Avg s ADC and Min s ADC than patients in DL group(all p values were less than 0.001).2.Among the clinical indicators,when take age > 50,post-menopause,abnormal vagina bleeding or discharge,WBC>6.37×109,NEUT>3.67×109,LYM<1.47×109,NLR>3.13 and SIRI>0.93 as respective standard to diagnose US+,the AUC range was 0.59-0.66;among qualitative MRI indicators,when take hyperintensity on DWI or hypointensity on delay phase as standard to diagnose US+,the AUC were0.89 and 0.69,respectively;among quantitative indicators,Avg ADC showed better performance than Min ADC;when take Avg w ADC<0.83 or Avg s ADC<0.70 as standard to diagnose US+,AUC were 0.75 and0.83,respectively.In consideration of the convenience and repeatability of measurement,we only included Avg w ADC for further analyses.3.The binary logistic regression demonstrated that,the predictive values of abnormal vagina bleeding or discharge,post-menopause,hyperintensity on DWI,hypointensity on delay phase,and Avg w ADC<0.83 were 2,1,5,2 and 2,respectively.When take total score >5 as standard to diagnose US+,the sensitivity,specificity,and AUC were 90.4,88.9 and0.96,respectively.4.In USE group,3 out of 32 histopathology proven US and u CS patients’ total score were equal or less than 5.The diagnose accordance rate of our scoring system in validation data was 90.6%.5.When take ≥5 points as cut-off value,the sensitivity,specificity and accuracy of diagnosing US+ were 95.6%,78.4% and 83.3%,respectively,and the sensitivity in USE group was 96.9%;when take ≥4 points as cut-off value,the sensitivity,specificity and accuracy of diagnosing US+were 98.2%,71.4% and 71.6%,respectively,and the sensitivity in USE group was 96.9%.Part three1.Patients in US+ groups were associated with higher age(p<0.01),rate of post-menopause status(p<0.01)and abnormal vagina bleeding or discharge(p < 0.01),maximum diameters of tumor(p=0.032),and average ADC value(p<0.01)than patients in DL group.2.The comparison of different convolutional neural networks(CNN of models based on deep-learning demonstrated that: Resnet50 and Xception of T2 model showed better performance;Inception V3 and Resnet50 of DWI model showed better performance;Resnet50 and Xception of mp-MRI model showed better performance.Hence,when to construct models based on image features to differentiate US+ and DL,Resnet50 might be more efficient.3.The comparison of different feature extract methods of complex mp models showed that: AUC range of models based on transfer-learning were higher than that of models based on radiomics.The comparison of5 evaluation indexes of 5 models demonstrated that,C-A model achieve the highest precision(0.90),and transfer-learning based complex mp model achieved the highest value in all the left 4 indexes.conclusion:1.In the differentiation of US and DL,when take CA125,or HE4,or LDH,or PLTL,or LYM,or NLR,or SIRI as individual indicator,the highest AUC of diagnosing US was 0.65.With the combination of above 7indicators,the AUC of diagnosing US was 0.72,with minor increase.Different indicators with significant difference could be found between three histopathological subtypes of US and DL groups,and this could help precise preoperative diagnosis.2.In the differentiation of US+ and DL,clinical and MRI indicators were included and a convenient corresponding clinic-imaging scoring model was built.When take >5 points as cut-off value could differentiate US+and DL with good performance(AUC:0.96).According to different clinical needs,different scoring threshold could be chosen to achieve better sensitivity or specificity.When this scoring model was used for external data validation,the diagnostic rate of US+ was 90.6%,showing good universality.3.Based on our data base,in the differentiation of US+ and DL,transferlearning of deep learning was feasible in radiomics and superior to radiomics.Machine learning model based on clinic-imaging-radiomics features have good diagnostic performance.