Exploratory Study Of Deep Learning In Magnetic Resonance Imaging Diagnosis Of Breast Cancer

Chapter 1 Weakly Supervised 3D Deep Learning for Breast Cancer Classification and Localization of the Lesions in MR ImagesPurpose:To evaluate the efficacy of 3D deep convolutional neural network(CNN)for diagnosing breast cancer and localizing the lesions at dynamic contrast enhanced(DCE)MRI data in a weakly-supervised manner.Materials and Methods:A retrospective study of breast cancer was conducted after approval by the institutional review board.The study included1537 breast dynamic enhanced MRI images and non-fat suppression images(mean age 47.5 years±11.8),including 1,529 women and 8 males,14 of whom had isolated lesions on both sides,a total of 1551 lesions.The research includes1033 lesions of malignant tumors and 518 lesions of benign lesions;time span from March 2013 to December 2017.All cases had pathological results of biopsy or surgery,and the imaging results were given a report of BI-RADS classification by a radiologist with 12 years of experience.Data preprocessing of breast segmentation was performed first,and Frangi method was applied to MRI 2D slices to obtain breast-air boundary,chest muscle boundary,and boundary between breast and fat.A series of morphological processing is performed,including threshold processing of the filtered slice,connected component analysis and cavity filling,and finally a binary mask of the breast region in each 2D slice is obtained.We then stacked all the sliced 2D masks and obtained a 3D breast segmentation mask for each MRI volume data.Next,we use a 3D Gaussian filter to smooth the breast mask.Using a 3D binary mask,we obtained a bounding box covering the entire breast area.Using a weakly supervised approach to localization of lesions in the breast,considering the features extracted in the early layers of the network with higher resolution,we recommend inferring the location of the lesion from the early layer rather than the last predictive layer.Aiming to effectively capture the volumetric spatial information of the MRI images,we employed the 3D CNN as infrastructure of our deep learning model.Specifically,we utilized a 3D DenseNet,i.e.,the layers in our network were densely connected,which encourages reuse of the features and helps improve the model performance.Under the supervision of medical image images,CNN is trained to achieve the location and benign and malignant classification of lesions.The entire data set was randomly divided into a training set(1073 cases),a validation set(157cases),and a test set(307 cases).The computational model diagnoses breast cancer with accuracy,sensitivity,specificity,and area under the receiver operating characteristic curve(AUC).Results:The location of breast cancer lesions is visually satisfactory.CNN improved the specificity while maintaining a relatively high cancer detection rate.In 307 verification data sets,there were 206 cases of pathological diagnosis of breast cancer,187 cases of deep learning convolutional neural network were correctly interpreted,19 cases were false negative;101 cases were confirmed by pathology.Cancer,CNN has 70 correct interpretations and 31 false positives.The accuracy rate of CNN diagnosis of breast cancer was 83.7%,the sensitivity was 90.8%,and the specificity was 69.3%.BI-RADS assessed the accuracy of breast cancer diagnosis as 85.7%,sensitivity was 98.5%,and specificity was59.4%.MRI is the most sensitive imaging modality for screening BC,The comparison between CNN and radiologist shows that the CNN model is comparable in accuracy to experienced radiologists,and its specificity is superior to radiologist,and the sensitivity is lower than the latter.Conclusion:The weakly-supervised learning method showed promise for localizing lesions in volumetric radiology images with only image-level labels.Deep learning with 3D neural networks demonstrated high performance for diagnosing breast cancer.Chapter 2 Prediction of Malignancy based on MRI Radiomics of Breast Cancer in a Visualized WayPurpose:To develop a Magnetic Resonance Imaging(MRI)radiomics based model for the malignancy prediction of breast cancer.Materials and Methods:Our study was conducted in a primary cohort of134 consecutive patients with mammographically suspicious findings.Breast MRI was done using a 1.5-T system(Magnetom Espree Pink;Siemens,Munich,Germany),equipped with an eight-channel breast coil.The model of self-organizing feature mapping(SOM)was used for data exploration and visualization.The decision tree was used to develop a prediction model.The performance of the proposed method on the cohort,the area under the curve(AUC)of receiver operating characteristic was assessed and compared with the classical BI-RADS Categories.Results:A cohort of 134 candidates with 49 diagnosed pathologically as benign and 85 as malignancy was studied.Five clinical candidate characteristics and 12 dominant lesion image-based ones were proposed to predict the malignancy of breast cancer.Through statistical analysis and data exploration,the selected 6 features including the ADC value,tumor size,margin definition,age,et al.were demonstrated to be significant predictors of the malignancy.The U-matrix generated by SOM showed a discrimination by appropriate segmentation intuitively.The heatmap of the U-matrix in the individual feature can give comprehensive understanding of the relationship between the malignancy and the separated feature.The proposed decision tree based model achieved differentiation of malignant from benign lesions with the AUC of 0.989with 7%of improvement over the traditional BI-RADS Assessment Categories based to predict the malignancy.Conclusion:Decision tree based model with only 6 features that performed well in the differentiation of malignant from benign lesions and achieved higher performance than BI-RADS based one.Chapter 3 Predicting Neoadjuvant Chemotherapy in Non-concentric Shrinkage Pattern of Breast Cancer Using ~1H Magnetic Resonance SpectroscopyPurpose:To explore the response to neoadjuvant chemotherapy(NAC)in non-concentric shrinkage pattern of breast cancer(BC)patients using ~1H magnetic resonance spectroscopy(MRS).Materials and Methods:Twenty-five BC patients were the study cohort.All patients received anthracycline-and taxane-based regimen as first-line treatment.Tumor response to chemotherapy was evaluated after the second and fourth cycles using magnetic resonance imaging and MRS.Final histopathology following surgery after 4–8 cycles of NAC served as a reference.Changes in total choline integral*(tCho)and tumor size in response versus non-response groups were compared using the two-way Mann–Whitney non-parametric test.Receiver operating characteristic(ROC)analyses were undertaken,and the area under the ROC curve(AUC)compared among them.Results:1H MRS revealed a negative tCho integral*in 6 cases at first follow-up and 14 cases at second follow-up.Based on pathology(Miller&Payne system),there were 16 cases of response,and 9 cases of non-response.The tCho integral*was significant difference between response and non-response group at the second follow-up(P=0.027).The tumor size changes were no significant differences in response and non-response groups at second follow-up study(P>0.05).The comparison of ROC curves among the change in tCho integral*and tumor size at baseline and both follow-ups revealed the maximum AUC of the change in tCho integral*to be 0.747 at second follow-up,sensitivity to be 93.75%,and positive predictive value(PPV)to be 78.9%.Conclusions:In non-concentric shrinkage pattern after NAC of breast cancer,when tumor size is difficult to reflect the response,tCho integral*reduction maybe a predict marker.