| Background: Colorectal cancer has the third worldwide incidence,nearly half of colorectal cancer patients may be diagnosed liver metastasis,which is also the main cause of death.Local treatment(e.g.surgical resection,ablation,etc.)is available for patients with colorectal cancer liver metastasis in clinical practice: for unresectable colorectal cancer liver metastasis that meet the indications for ablation treatment,ablation can improve their overall survival;for complete resection of colorectal cancer liver metastasis,the median survival time was 35 months,and the 5-year survival rate was 30%-57%.However,patients with ablation could have risk of local tumor progression,and patients with surgical resection also have a certain risk of recurrence.Therefore,evaluating the risk of local tumor progression and recurrence after local treatment of liver metastasis is essential to predict patient prognosis.Radiomics,as a new method of image analysis,could diagnose diseases by a non-invasive method and predict the outcome of lesions and treatment response.Deep learning plays a role in the classification,detection,segmentation and image generation,and also has certain adaptive learning functions.Pathomics enables the diagnosis,prognosis assessment and treatment response prediction for diseases.Previous studies have found that before and after ablation CT-based radiomics models could predict the local tumor progression of patients with colorectal cancer liver metastasis,and CT-based radiomics also plays a certain role in predicting the total survival of patients with colorectal cancer liver metastasis.The combination of CT-based radiomics and deep learning model has high diagnostic efficacy and predictive efficacy on the response of solid tumor after chemotherapy.The combination multi-omics model of CT-based radiomics,deep learning and pathomics has good predictive efficacy for the overall survival of colorectal cancer lung metastasis.However,there were few researches focusing on the colorectal cancer liver metastasis MRI related radiomics and deep learning model after ablation,and there were also few researches on MRI deep learning combined with pathomics after hepatectomy.Therefore,the present study intends to predict local tumor progression of colorectal cancer liver metastasis based on the machine learning model of MRI-based radiomics before ablation treatment and to evaluate the model effectiveness.To predict local tumor progression in the ablation treatment of colorectal cancer liver metastasis based on MRI-based radiomics and deep learning model of the ablation region edge after ablation treatment,combined with clinical parameters,and to evaluate the model effectiveness.To predict the prognosis,evaluate model effectiveness and risk stratification of local resection of colorectal cancer liver metastasis based on MRI-based deep learning and pathomics,combined with clinical parameters.Methods:Part I: Clinical study on the application of pre-ablation MRI-based radiomics to predict local tumor progression for colorectal cancer liver metastasis.From January 2015 to January 2022,patients with colorectal cancer liver metastasis who underwent local ablation in the First Affiliated Hospital of China Medical University were included,including 94 pre-ablation liver lesions.Hepatic MRI dynamic contrast-enhanced scan(T1WI,T2 WI and VP sequences)before ablation were obtained through PACS,clinical information was collected by electronic medical records,and the local tumor progression after local ablation was collected by checking radiology images.After image preprocessing(resample,N4 bias field correction),two radiologists with more than 5 years of experience delineate the ROI of liver metastasis manually by ITK-SNAP software.Pyradiomics was performed for feature extraction.Then the features were screened through regularization,U-test and Pearson correlation analysis.The best λ value was calculated through LASSO for feature dimension reduction.The radiomics model was constructed by using the best features.ROC,DCA decision curve and confusion matrix were used for predicting the model effectiveness.Construction of a nomogram for predicting local tumor progression using machine learning model radiomics signature in combination with clinical parameters and evaluation of their efficacy.Part II: Clinical study on the application of post-ablation MRI-based radiomics,deep learning and clinical multimodal to predict local tumor progression for colorectal cancer liver metastasis.From January 2015 to January 2022,patients with colorectal cancer liver metastasis who underwent local ablation in the First Affiliated Hospital of China Medical University were included,including 82 post-ablation liver lesions.Hepatic MRI dynamic contrast-enhanced scan(T1WI,T2 WI and VP sequences)after ablation were obtained through PACS,clinical information was collected by electronic medical records,and the local tumor progression after local ablation was collected by checking postoperative radiology images of the patient.After image preprocessing(resample,N4 bias field correction),two radiologists with more than 5 years of experience delineate the ROI of liver metastases manually by ITK-SNAP software.The post ablation area of the whole lesion was sketched manually,while the ablation area edge(ablation edge area with 4voxels inward and 6 voxels outward)and the edge area layering(the width of each layer is 2 voxels)of the post-ablation lesion were sketched both manually and automatically.For the feature extraction of ablation area edge,pyradiomics was performed for radiomics feature extraction,and deep learning convolution neural network(CNN)model was performed for deep learning feature extraction.For the feature extraction of ablation area edge layer,pyradiomics was performed for radiomics feature extraction.Then the features were screened through regularization,U-test and Spearman correlation analysis.The best λ value was calculated through LASSO for feature dimension reduction.The MRI-based radiomics combined with deep learning model was constructed by using the best features.ROC,DCA decision curve and confusion matrix were used for predicting the combined model effectiveness and clinical decision-making.Clinical machine learning models were constructed by applying clinical parameters to obtain clinical model signatures,combing with radiomics-deep learning signatures for drawing nomogram and predicting the risk of local tumor progression.Part III: Clinical study on the application of deep learning multi-omics model to predict the recurrence risk of surgical resection for colorectal cancer liver metastasis.From January 2015 to January 2022,patients with colorectal cancer liver metastasis who underwent liver resection surgery in the First Affiliated Hospital of China Medical University were included,including 107 patients.Hepatic MRI dynamic contrast-enhanced scan(T1WI,T2 WI and VP sequences)before resection were obtained through PACS,clinical information was collected by electronic medical records,and the recurrence after local resection was collected by checking radiomic images.After image preprocessing(resample,N4 bias field correction),two radiologists with more than 5years of experience delineate the ROI of liver metastases manually by ITK-SNAP software.Radiomic deep learning features were extracted by deep learning convolution neural network(CNN)model.For time-dependent MR-based deep learning model,LASSO-COX was used for feature screening and model construction to obtain deep learning signature associated with PFS.A nomogram was constructed based on deep learning signatures and clinical parameters for risk stratification and point-in-time recurrence rate prediction and efficacy assessment of surgically resected patients.The tumor area of WSIs was delineated manually at 40 × by Qu Path software(version 0.3.2),while pathomic deep learning features were extracted by Vgg19.Then the combined features were screened through regularization,U-test and Spearman correlation analysis.The best λ value was calculated through LASSO for feature dimension reduction.The combined model of radiomics deep learning and pathomic was constructed by using the features.ROC and DCA decision curve were used for predicting the combined model effectiveness.Clinical parameters,after the univariate analysis,combing with the model calculated radio-pathomic deep learning signatures,were performed for drawing nomogram and predicting the risk of recurrence.Results:Part I: Clinical study on the application of pre-ablation MRI-based radiomics to predict local tumor progression of colorectal cancer liver metastasis.A total of 3591 features from manually sketched ROI were extracted through pyradiomics,including 1197 features each extracted from T1 WI,T2WI and VP sequences.According to the local tumor progression after ablation treatment,patients were divided into two groups: 35 in the progressive group and 59 in the non-progressive group.After U-test,Pearson correlation analysis and Lasso model dimension reduction,21 best features(10 T1 WI features,5 T2 WI features,and 6 VP features)were selected and used to construct the MRI radiomics model through the 5-fold cross-validation and different types of machine learning algorithms(LR,SVM,KNN,Random Forest,Extra Trees,XGBoost,Light GBM,MLP).Through evaluating the effectiveness of the model,the AUC of the LR model in training set was 0.92(95% confidence interval0.86-0.98),the AUC of the validation set was 0.86(95% confidence interval 0.68-1.00),the prediction accuracy of the model was 0.74,the sensitivity was 0.67,and the specificity was 0.77.The area under the DCA curve suggests that the prediction of MRI-based radiomics model may benefit the patients with colorectal cancer liver metastasis.A nomogram for predicting local tumor progression was constructed by applying MRI-based radiomics signature in combination with clinical parameters,but the model efficacy was not further improved compared to the radiomics model alone.Part II: Clinical study on the application of post-ablation MRI-based radiomics,deep learning and clinical multimodal to predict local tumor progression for colorectal cancer liver metastasis.Pyradiomics was performed for the radiomics feature extraction of ablation area edge and ablation area edge layering.Deep learning convolution neural network(CNN)model was performed for the deep learning feature extraction of ablation area edge.According to the local tumor progression after ablation treatment,patients were divided into two groups:33 in the progressive group and 49 in the non-progressive group.After U-test,Spearman correlation analysis and Lasso model dimension reduction,9 best features were extracted from the edge of the ablation area(3 T1 WI image features,3 T1 deep learning features,1T2 WI image features,2 VP image features),5 best features were extracted and dimension reduced from the layer III of the ablation area edge(1 T1 WI feature,3 T2 WI features,and 1 VP feature).The best 9 features at the edge of the ablation area were used to construct the combing model of MRI radiomics and deep learning through the 5-fold cross-validation and different types of machine learning algorithms.After evaluating the effectiveness of the model,the AUC of LR model,SVM model,KNN model in both training set and validation set can reach more than 0.90,the accuracy is higher than 0.70,and the sensitivity of validation set is high,suggesting that the combined model of MRI-based radiomics and deep learning may benefit patients with colorectal cancer liver metastasis ablation treatment.The nomogram of MRI-based radiomics-deep learning signature and clinical KNN model signature has higher predictive efficacy than clinical signature or radiomics-deep learning signature in both the training set(AUC=0.97,95%CI: 0.94-1.00)and test set(AUC=0.98,95% CI: 0.95-1.00).The model was well calibrated and was useful in assessing local tumor progression.Further stratification of the ablation zone edges,the five best features of region III were used for radiomics model construction,and the KNN model had an AUC of 0.77(95% CI: 0.5-1.0),an accuracy of 0.76,and a sensitivity and specificity of 0.60 and 0.91.It is suggested that the medial margin region of ablation may play an important role in the prediction of local tumor progression.Part III: Clinical study on the application of deep learning multi-omics model to predict the recurrence risk of surgical resection for colorectal cancer liver metastasis.The deep learning model was used to extract deep learning features,and the MRI-based deep learning model was constructed after feature screening with LASSO-COX to obtain point-in-time dependent deep learning model signature,which had good predictive efficacy(AUC>0.70)and good calibration for predicting recurrence rate at 6,12 and 18 months time points when combined with clinical parameters,and the nomogram could predict the recurrence risk of patients at different time points.The results of Kaplan-Meier survival analysis suggested that compared to low-risk patients,PFS is lower in high-risk patients,and the difference was statistically significant in both the training and test sets(P<0.05).According to the recurrence after local resection treatment,patients were divided into two groups: 27 recurred early after 6 months of surgical treatment,while 80 did not;51 cases recurred after 12 months of surgical treatment,while 56 did not.Radiomic deep learning features and pathomics deep learning features were extracted by CNN model.After U-test,Spearman correlation analysis and Lasso model dimension reduction,19 best features were extracted in the 6-month recurrence group(7 T2 WI deep learning features,9 VP deep learning features,and 3 pathomics deep learning features),while 19 best features in the 12-month recurrence group(7 T2 WI deep learning features,10 VP deep learning features,and 2 pathomics deep learning features).The best features were used to construct the combing model of MRI deep learning and pathomics deep learning model through the 5-fold cross-validation and different types of machine learning algorithms.Then the model effectiveness in 6-month recurrence group was evaluated with a relative high AUC(>0.70)in both SVM and LR model,a slightly higher accuracy(0.82)and up to 100% sensitivity in the SVM model.There was no significant statistical difference among the models(P>0.05)by Delong test.In the 12-month recurrence group,the AUC of SVM,LR and KNN models can reach more than 0.80.The nomogram constructed by radiomic-pathomic deep learning signatures and clinical parameters(serum CA199 level,targeted treatment acceptance,liver metastasis type,lymph node metastasis in the 6-month group;age,serum CA199 level,targeted treatment acceptance,liver metastasis type in the 12-month group)can predict the 6 month and 12 month recurrence rate.Conclusions: Pre-ablation MRI-based radiomics model has a good predictive effect on the local tumor progression of colorectal cancer liver metastasis.Post-ablation MRI-based radiomics-deep learning combined model has a good predictive effect and high sensitivity on the local tumor progression of colorectal cancer liver metastasis.The combined MRI-based radiomics,deep learning and clinical(tumor size,serum CEA level,serum CA199 level)model is better than the single-omics model in predicting local tumor progression after ablation,playing a key role in predicting the risk of local tumor progression,which can provide reference for further examination and clinical decision.The MRI-based radiomics model constructed by edge stratification of the ablation region suggests that the medial edge region of ablation may play an important role in the prediction of local tumor progression.MRI-based deep learning signature combined with clinical parameters can predict patients’ risk of recurrence at different time points and stratify the population recurrence risk for patients with colorectal cancer liver metastasis.More aggressive screening measures and treatments can be considered for patients in the high-risk group.The SVM machine learning model in radiomics-pathomics deep learning model has high predictive efficacy(AUC>0.70,accuracy>0.70)and high sensitivity for the prognosis(6-month recurrence,12-month recurrence)of patients with surgically resected colorectal cancer liver metastasis,which can provide a reference for clinical decision-making.Nomogram constructed by radiomics-pathomics signature combined with clinical parameters was able to predict recurrence of colorectal cancer liver metastasis at 6 months and 12 months after surgical resection. |
PDF Full Text Request