| Objective: A large number of studies have found that Ki-67 is an independent prognostic factor for gastrointestinal stromal tumors,which is highly correlated with the indicators of NIH risk classification(2008 improved version)of gastrointestinal stromal tumors,and there is an obvious correspondence in the risk classification,which is expected to become a supplement.However,results varied widely between different observers.At present,when studying the relationship between Ki-67 and GIST prognostic indicators,the counting methods of various types of Ki-67 are not uniform,and the critical value obtained is poorly repeated,which limits the predictive value of Ki-67 for prognosis,and Ki-67 is not included in the reference index of risk classification in the NIH risk classification of gastrointestinal stromal tumors(2017improved version).This experiment aims to explore a new widely shared Ki-67 index method that minimizes the influence of human subjectivity,and to provide an experimental basis for further exploring the prognostic value of Ki-67 for gastrointestinal stromal tumors.Method: A total of 48 patients with surgical resection of GIST from January 2015 to January 2022 at the First Affiliated Hospital of Dali University were selected as the study subjects.Under the Image J platform,the "image feature classifier"(machine automatic counting)was trained to analyze the Ki-67 positivity of GIST tumor cells and count.The overlap of the three interpretation methods was verified by comparing the Ki-67 index readings of the hot spot region framing manual estimation method under digital slices and the artificial microscopic estimation method.Repeatability was explored by comparing the within-group correlation coefficient ICC for calculating the "machine automatic counting Ki-67 index" with the manually estimated reading.The Ki-67 value obtained by the trained "image feature classfier" was used to explore the relationship with clinicopathological features,and the corresponding model of GIST risk classification(NIH2008 improved version)was constructed.Result: 1.After the training set(7 cases),the number of events where the interpretation results of the "Image Feature Classifier" were inconsistent with the manual interpretation continued to decrease,and the effect was significant when 4cases were added as the training set.2.The interpretation results of the "picture feature classifier" are consistent with the results of the manual counting method of framing hot spots under digital slices and the estimation method under artificial microscope,and the consistency of the first two is excellent.3.Reproducibility study of machine evaluation and manual evaluation: the ICC coefficient of Ki-67 manually evaluated by three physicians in the hot spot area under the microscope was 0.768(95% confidence interval 0.621,0.836),and the ICC coefficient of training three "picture feature classifiers" to evaluate Ki-67 was 0.974(95% confidence interval0.958,0.984),which can be considered to be better repeatability of manual evaluation of hot spot area under microscope and hot spot area counting of manual training machine.Moreover,the reproducibility of training "image feature classifier" to evaluate Ki-67 is better than that of manual evaluation under the microscope of hot spots.Conclusion: The Ki-67 value estimated by the training machine under the Image J platform can well correspond to the risk grading of gastrointestinal stromal tumor,and the reproducibility is excellent,which has unique advantages compared with manual evaluation under the microscope.The resulting critical values for Ki-67 were 2.2%,4.7%.It can be recommended as one of the prognostic indicators. |
PDF Full Text Request