Prediction Of Histological Grade And Ki-67 Expression In Breast Cancer Based On Multi-Parameter Imaging And Multi-task Learning

The incidence of breast cancer is the highest in women's cancer,and with high mortality rate.With the increasing incidence and the rejuvenation of the patient population,so it is more and more important to provide accurate treatment plans based on prognostic indicators.Magnetic resonance imaging(MRI)is a multi-parameter imaging in which the complementarity of information between different parametric imaging techniques makes MRI examination more conducive to cancer diagnosis.Dynamic contrast enhancement MRI(DCE-MRI)can not only provide morphological information of tumors,but also microscopic information related to tissue blood flow.And diffusion weighted imaging(DWI)can provide qualitative and quantitative information about diffusion characteristics,so it adds a new dimension for MRI examination.The histopathological grade of breast cancer is closely related to its prognosis.Ki-67 is an important biomarker,and both of these indicators provide important guidance for the accurate diagnosis and treatment of breast cancer.Multi-task learning improves the learning effect by considering the connections between different tasks and learning different tasks.It has not been studied in breast cancer.The study of histopathological grading of breast cancer is a relatively advanced field of exploration,and similar studies rarely predicted this aspect.This study predicted the histological grade and Ki-67 expression of breast cancer by multi-task learning methods combined with multi-parametric magnetic resonance(DCE-MRI and DWI)images.The content of this paper has the following points:1)Pathological information processing:We analyze 144 cases which have MRI examination and their BI-RADS grade is 3 or above 3.The cases that we adopt conform the demands of our research.And the pathological information includes histopathological grading,value of Ki-67 and age.Then statistical analysis was carried out on the summarized information by means of analysis of variance and chi-square test.According that,we can diagnosis the correlation between Ki-67 and histopathological grading.And judge other information has no effect on the expression of Ki-67.2)Segmentation and image feature acquisition of multi-parametric image lesions:For DCE-MRI images,we used spatially blurred C-means and Markov random field algorithms to obtain three-dimensional lesions.We calculated the ADC map based on the b value of the DWI image,then registered the lesion region of the above-mentioned segmented DCE-MRI sequence with the corresponding ADC image,and finally the lesion area consistent with the contour of the DCE-MRI lesion was acquired on the ADC image.We then extracted features from the acquired lesion images of each sequence.3)Histological grading and Ki-67 prediction of single-task learning studies:We performed single-feature analysis,single-parameter modeling analysis,and multi-parameter model fusion analysis for two prognostic indicators.We use logistic regression in the single feature analysis and.Two methods that Lasso with logistic regression and SVM-RFE with SVM are applied in the single parameter modeling analysis.As for multi-parameter model fusion,we use multi-classifier fusion.Mentioned above we calculate AUC,sensitivity and specificity to evaluate prediction model.4)Histological grading and Ki-67 prediction of multi-task learning studies:According to the results of the chi-square test,we found that the correlation between histological grade and Ki-67 was significant(P<0.05).In order to improve the learning performance of single task,we choose the multi-task learning method based on feature sharing(ie multi-task feature selection)to learn the two prediction tasks at the same time.After multitasking feature selection,we use the IND method to model predictive metrics under a general convex framework for multitasking learning.Single parametric images train each multi-task prediction models.Like single-task learning,we use multi-classifier fusion to fuse prediction models with different parameters to compensate for the shortcomings of single-parameter prediction models.We have comparatively analyzed the predictive power of different learning methods at the end.This paper study breast cancer grading and Ki-67 expression prediction based on multi-parameter imaging and multi-task learning methods.The results show that in single-task learning,the optimal results of the multi-parameter joint application of the two prediction tasks are higher than the single-parameter prediction model.However,in multi-task learning,the optimal results predicted on different parametric images are better than single-task learning.The combination of multi-parametric imaging and multi-task learning methods resulted in an optimal AUC of 0.816±0.072 for the hierarchical prediction and an optimal AUC of 0.821 ± 0.074 for the Ki-67 expression prediction.The results showed that breast cancer analysis based on multi-parameter imaging and multi-task learning methods can significantly improve the histological grade and the predictive ability of Ki-67 expression,and provide a more accurate basis for clinical diagnosis and prognosis.