Prediction model training method for predicting relative curative effect of EC and T in NAC

Abstract

The invention discloses a prediction model training method for predicting the relative curative effect of EC and T in NAC, and the method comprises the steps: building a data set through obtaining the MRI image data and clinical pathological data of a subject, and building a radiomics model based on DCE and ADC sequences through the steps of image segmentation, feature extraction, qualitative relative curative effect features, model building and verification, and the like. And predicting the relative curative effect of EC and T treatment in combination with a mixed model of clinical pathological factors. According to the method, the relative curative effect of EC (epirubicin and cyclophosphamide) and T (taxus drugs) treatment schemes can be effectively predicted in the middle stage of NAC, so that the treatment schemes are timely adjusted, and the treatment effect is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of neoadjuvant chemotherapy, and specifically to a method for training a prediction model for predicting the relative efficacy of EC and T in NAC. Background Art

[0002] Neoadjuvant Chemotherapy (NAC) has become an important part of breast cancer treatment. Its main purpose is to reduce the tumor volume and improve the success rate of breast-conserving surgery. However, there are significant individual differences in the responses of patients to the same treatment regimen during NAC, which are not only reflected among different patients but also in different treatment stages of the same patient. At the same time, currently, only after the completion of NAC treatment can complete tumor volume change rate data be obtained to evaluate the treatment effect, but at this time, the treatment regimen cannot be adjusted anymore. Therefore, accurately predicting the relative efficacy of different treatment regimens in NAC is of great significance for timely adjusting the treatment regimen and optimizing personalized treatment strategies.

[0003] In recent years, with the continuous development of imaging technology, especially the wide application of magnetic resonance imaging (MRI) technology, it has provided a more accurate tool for evaluating tumor response during NAC. MRI can not only observe the size and volume changes of tumors but also provide detailed information on tumors and their microenvironments through multi-parameter imaging techniques. The radiomics method based on MRI extracts quantitative information from medical images, reveals the association between medical images and pathophysiological information, and provides a new way for optimizing clinical diagnosis and treatment strategies.

[0004] Currently, the research on using radiomics to predict the relative efficacy of different treatment regimens in NAC is gradually deepening. Researchers have tried to predict the relative efficacy of the EC (epirubicin plus cyclophosphamide) and T (taxane drugs) treatment regimens during the middle stage of NAC by constructing a machine learning model based on MRI radiomics, with the expectation of timely adjusting the treatment regimen and improving the treatment effect. Summary of the Invention

[0005] The purpose of this part is to outline some aspects of the embodiments of the present invention and briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this part, as well as in the abstract and title of the specification of this application, to avoid obscuring the purpose of this part, the abstract, and the title. However, such simplifications or omissions cannot be used to limit the scope of the present invention.

[0006] Therefore, the purpose of the present invention is to provide a method for training a prediction model for predicting the relative efficacy of EC and T in NAC based on MRI radiomics and clinicopathology, so as to solve the problem that the existing NAC treatment regimen cannot predict the relative efficacy and thus cannot timely adjust the treatment method, which affects the treatment effect.

[0007] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0008] A method for training a prediction model for predicting the relative efficacy of EC and T in NAC, the steps are as follows:

[0009] S1. Screen clinical data: Obtain the imaging data and clinical pathological information of the subjects to construct a data set;

[0010] S2. Image segmentation: Select the second post-contrast phase image of the DCE sequence for segmentation among 6 dynamic phases. Manually depict the region of interest (ROI) containing the entire MRI-visible tumor layer by layer on the axial plane, and then expand the intra-tumor region isotropically in three dimensions by 5 mm to obtain the peritumoral region;

[0011] S3. Feature extraction: Extract radiomics features from the images. Before feature extraction, standardize the intensity distribution of the images, and then extract relevant features from each region of interest of the DCE sequence and the apparent diffusion coefficient map, and calculate the difference between the radiological feature values of the NAC-pre and NAC-mid images as the Delta radiomics feature to capture the longitudinal changes of the tumor;

[0012] S4. Features of qualitative relative efficacy: Use the tumor volume change rate to characterize the relative efficacy of the first 4 EC treatment cycles and the subsequent 4 T treatment cycles;

[0013] S5. Construct a radiomics model: The model construction is divided into two parts: The first part evaluates the performance of the model using multiple sequences and multiple regions of interest through ablation experiments. The second part integrates radiomics features in various aspects to construct the Original model, Delta model, and Fusion model to predict the relative efficacy of EC and T treatments using multiple sequences and multiple regions of interest;

[0014] S6. Construct a hybrid model: The outputs of the Original, Delta, and Fusion radiomics models for each patient are used as radiomics features. Using the logistic regression method, combine the radiomics features with important clinical pathological factors to establish a hybrid model;

[0015] S7. Statistical analysis and evaluation: Evaluate the performance of the radiomics model and the hybrid model.

[0016] As a preferred embodiment of the prediction model training method for predicting the relative efficacy of EC and T in NAC of the present invention, in step S1, the imaging data is MRI imaging data, including imaging data at three key points: NAC-early stage, NAC-middle stage, and NAC-late stage. The imaging data includes at least one localization sequence, axial TIRM sequence, DWI, fat-suppressed DCE sequence, and sagittal fat-suppressed T2-weighted imaging.

[0017] As a preferred embodiment of the prediction model training method for predicting the relative efficacy of EC and T in NAC of the present invention, in step S1, the clinicopathological data includes gender, age, menstrual status, clinical T stage, clinical N stage, ER status, PR status, HER2 status, and Ki67.

[0018] As a preferred embodiment of the prediction model training method for predicting the relative efficacy of EC and T in NAC of the present invention, in step S2, use 3D Slicer software to manually delineate the region of interest (ROI) containing the entire MRI-visible tumor layer by layer in the axial plane.

[0019] As a preferred embodiment of the prediction model training method for predicting the relative efficacy of EC and T in NAC of the present invention, in step S2, use the SimpleITK package in Python to isotropically expand the intratumoral region by 5 mm in three dimensions.

[0020] As a preferred embodiment of the prediction model training method for predicting the relative efficacy of EC and T in NAC of the present invention, in step S3, use PyRadiomics to extract radiomics features from the images.

[0021] As a preferred embodiment of the prediction model training method for predicting the relative efficacy of EC and T in NAC of the present invention, in step S4, the relative efficacy is divided into two types: the first is to determine the efficacy of 4 EC treatment cycles by calculating the relative net change in tumor volume from the NAC-early stage to the NAC-middle stage; the second is to calculate the relative net change from the NAC-middle stage to the NAC-late stage to determine the efficacy of 4 T treatment cycles, and then evaluate the relative efficacy of the treatment by comparing these net change values.

[0022] As a preferred embodiment of the prediction model training method for predicting the relative efficacy of EC and T in NAC of the present invention, in step S5, multiple sequences and multiple regions of interest mainly use two MRI sequences: DCE and ADC, and two regions of interest: intratumoral region and peritumoral region.

[0023] As a preferred solution of a prediction model training method for predicting the relative efficacy of EC and T in NAC according to the present invention, in step S5, feature selection includes the following steps:

[0024] Select features with a variance greater than 1.0 through a variance threshold;

[0025] Use the Mann-Whitney U test to select features related to the relative efficacy of NAC treatment;

[0026] Use the random forest model to rank the feature importance and select the top 100 most important features;

[0027] Use the least absolute shrinkage and selection operator regression and 10-fold cross-validation to select features with non-zero coefficients.

[0028] As a preferred solution of a prediction model training method for predicting the relative efficacy of EC and T in NAC according to the present invention, in step S5, the Original model combines the radiomics features of NAC-pre and NAC-mid, the Delta model includes the Delta radiomics features changing from NAC-pre to NAC-mid, and the Fusion model combines all the above features.

[0029] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention constructs a radiomics model based on DCE and ADC sequences and a hybrid model combined with clinicopathological factors based on MRI radiomics, and finally obtains the Delta model of FusionOriginal and changing radiomics features, as well as the Delta+clinicopath hybrid model combined with clinicopathological factors, which shows the best performance in predicting the relative efficacy of EC and T treatments; this method can effectively predict the relative efficacy of EC (epirubicin plus cyclophosphamide) and T (taxane drugs) treatment regimens in the mid-term of NAC, so as to timely adjust the treatment plan and improve the treatment effect. Brief Description of the Drawings

[0030] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the present invention will be described in detail below with reference to the drawings and detailed embodiments. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts. Among them:

[0031] Figure 1 It is the model training flowchart of the prediction model training method for predicting the relative efficacy of EC and T in NAC based on MRI radiomics and clinicopathology provided by the embodiment of the present invention;

[0032] Figure 2 The flowchart of patient recruitment provided by the embodiment of the present invention;

[0033] Figure 3 The ROC curve graph of the Delta model in the training set provided by the embodiment of the present invention;

[0034] Figure 4 The ROC curve of the Delta model in the test set provided by the embodiment of the present invention;

[0035] Figure 5 The calibration curve of the Delta model in the training set provided by the embodiment of the present invention;

[0036] Figure 6 The calibration curve of the Delta model in the test set provided by the embodiment of the present invention;

[0037] Figure 7 The DCA curve of the Delta model in the training set provided by the embodiment of the present invention;

[0038] Figure 8 The DCA curve of the Delta model in the test set provided by the embodiment of the present invention. Detailed implementation manners

[0039] To make the above objects, features, and advantages of the present invention more obvious and understandable, the following will describe the detailed implementation manners of the present invention in conjunction with the accompanying drawings.

[0040] The present invention provides a prediction model training method for predicting the relative efficacy of EC and T in NAC based on MRI radiomics and clinical pathology, which solves the problem that the existing NAC treatment plan cannot predict the relative efficacy and thus cannot adjust the treatment method in time, affecting the treatment effect. As Figure 1 shown, the specific scheme is as follows:

[0041] Step 1: Screen patients with complete MRI images and clinical pathology data. As Figure 2 shown in the recruitment process, female patients diagnosed with breast cancer from April 2016 to March 2023 were included. The eligibility criteria were: (1) invasive breast cancer; (2) adopting the T regimen after completing all EC cycles; (3) MRI data obtained before NAC treatment (NAC-pre), before the fifth cycle (NAC-middle), and before surgery (NAC-late); (4) complete clinical and pathological data. The exclusion criteria were: (1) incomplete or non-standard NAC treatment; (2) insufficient MRI quality or lack of MRI data; (3) previous chemotherapy or targeted therapy; (4) distant metastatic lesions.

[0042] Collect the clinical data of each patient before NAC treatment, including gender, age, menstrual status, clinical T stage and clinical N stage. The patients were divided into a training set and a test set at a ratio of 7:3.

[0043] A total of 303 patients were excluded from this study, including those with insufficient MRI data (n = 292) and insufficient image quality (n = 11). Therefore, 190 patients were included in this study. Table 1 summarizes the baseline characteristics of all patients. In the training set, the proportion of patients with EC superior to T was 48.1% (64:133), and in the test set it was 47.4% (27:57). Among the patients in both groups, there were no significant differences in the clinicopathological characteristics of patients with EC superior to T and T superior to EC (all p values > 0.05).

[0044] Table 1 - Patient baseline characteristics

[0045]

[0046]

[0047] The clinicopathological and treatment strategies were as follows: All patients received 4 cycles of EC (epirubicin, 90 mg / m 2 , on day 1, every 14 or 21 days; cyclophosphamide, 600 mg / m 2 , on day 1, every 14 or 21 days), followed by 4 cycles of taxane treatment. Taxane treatment included albumin-bound paclitaxel (Abraxane, 260 mg / m 2 , on day 1, every 14 days), solvent-based paclitaxel (paclitaxel, 175 mg / m 2 , on day 1, every 14 days), docetaxel, 75 mg / m 2, on day 1, every 21 days) or liposome (paclitaxel liposome injection, 175 mg / m2, on day 1). For patients with epidermal growth factor receptor 2 (HER2)+, starting from the 5th cycle, trastuzumab (Herceptin, 6 mg / kg every 21 days, 8 mg / kg loading dose) was added. Before each NAC cycle, the patient's complete blood count, liver function, renal function, and electrocardiogram were monitored. If the patient developed febrile neutropenia, grade 4 neutropenia, grade 4 thrombocytopenia, or grade 4 non-hematological toxicity (excluding nausea, vomiting, and fatigue), the dose of the NAC regimen was reduced. Immunohistochemistry (IHC) was used to detect the status of estrogen receptor (ER), progesterone receptor (PR), HER2, and Ki67. If the tumor showed 1% nuclear staining cells, it was classified as ER / PR positive. The HER2 status was evaluated as negative (HER2-), IHC grade 0 and 1+, positive (HER2+), IHC grade 3+. For tumors with IHC grade 2+, fluorescence in situ hybridization (FISH) was used to detect HER2 gene amplification. The expression of Ki-67 was evaluated using a cut-off index of 30%; an expression below 30% was considered low, while ≥30% was considered high.

[0048] The MRI image acquisition process and protocol were as follows: MRI scans were performed in the prone position using a 1.5T or 3.0T magnetic resonance at three key points: pre-NAC, mid-NAC, and post-NAC. The protocol included at least one localization sequence, axial TIRM sequence, DWI, fat-suppressed DCE sequence, and sagittal fat-suppressed T2-weighted imaging (T2WI). The dose of gadolinium-DTPA was 0.1 mmol / kg, and the infusion rate was 3 mL / s, followed by a 20 mL saline flush. ADC maps were generated from the DWI images using two b-values.

[0049] The 1.5T magnetic resonance imaging protocol included: (1) axial TIRM (TR / TE: 3450 / 61 ms; field of view: 340 mm × 340 mm; matrix: 320 × 320; flip angle: 80°; slice thickness: 4 mm); (2) axial DWI (b-value: 50 / 800 s / mm 2 ; TR / TE: 5200 / 65 ms; field of view: 323 mm × 161 mm; slice thickness: 5 mm); (3) DCE sequence (TR / TE: 4.23 / 1.57 ms; field of view: 340 mm × 340 mm; matrix: 448 × 448; slice thickness: 1 mm; flip angle: 10°; pixel resolution: 0.8 × 0.8 × 1.0 mm 3; Temporal resolution: 1 minute), repeated 5 times before and after administration; (4) Sagittal fat-suppressed T2WI (TR / TE: 2910 / 72 ms; Field of view: 200 mm × 200 mm; Matrix: 320×320; Flip angle: 80°; Slice thickness: 4 mm). The total scan time was 16 minutes and 55 seconds.

[0050] 3.0T magnetic resonance imaging, imaging protocol includes: (1) Axial TIRM (TR / TE: 5320 / 57 ms; Field of view: 340 mm × 340 mm; Matrix: 384×384; Flip angle: 150°; Slice thickness: 4 mm); (2) Axial DWI (b value: 50 / 800 s / mm 2 ; TR / TE: 7500 / 64 ms; Field of view: 350 mm × 163 mm; Matrix: 180×84; Flip angle: 180°; Slice thickness: 5 mm); (3) DCE sequence (TR / TE / TE: 3.90 / 1.66 ms; Field of view: 360 mm × 360 mm; Matrix: 320×320; Flip angle: 10°; Slice thickness: 1.5 mm; Voxel resolution: 1.1×1.1×1.5 mm3; Temporal resolution: 1 minute), repeated 5 times before and after administration; (4) Sagittal fat-suppressed T2WI (TR / TE: 4000 / 70 ms; Field of view: 250 mm × 250 mm; Matrix: 320×320; Flip angle: 150°; Slice thickness: 4 mm). The total scan time was 18 minutes and 6 seconds.

[0051] Step 2: MRI image cutting and feature extraction: Select the second post-contrast phase image of the DCE sequence for segmentation among 6 dynamic phases. Manually delineate the region of interest (ROI) containing the entire MRI-visible tumor layer by layer in the axial plane using 3D Slicer software. Then, use the SimpleITK package in Python 3.9.13 to expand the intratumoral region isotropically by 5 mm in three dimensions to obtain the peritumoral region. A radiologist with 5 years of breast imaging experience segmented all cases. To ensure inter-observer consistency, another radiologist with 5 years of experience re-segmented 50 randomly selected cases. Both radiologists were blinded to the clinical and histopathological data.

[0052] Radiomics features were extracted from the Original and filtered images using PyRadiomics. Before feature extraction, the intensity distribution of the images was normalized. A total of 1218 features were extracted from each ROI of the DCE sequence, and 1132 features were extracted from the ADC map. Considering two ROIs (intratumoral region and peritumoral region), the imaging data of the pre-NAC and during-NAC images for each patient contributed a total of 9400 features from the DCE and ADC images. To capture the longitudinal changes of the tumor, we calculated the delta radiomics features as the differences between the radiomics feature values of the pre-NAC and during-NAC images. Therefore, a total of 14,100 radiomics features were generated for each patient.

[0053] Step 3: Integrate specific MRI sequences and ROI features to preliminarily construct multiple radiomics models: The radiomics models were mainly constructed using two MRI sequences (DCE and ADC) and two ROIs (intratumoral region and peritumoral region). To evaluate the inter-observer reproducibility, we calculated the intraclass correlation coefficient (ICC) for each radiomics feature. Only the features with satisfactory inter-observer reproducibility, defined as ICC ≥ 0.80, were retained in the model.

[0054] For each radiomics model, the following feature selection process was performed in four steps: (1) Select features with a variance greater than 1.0 by variance threshold; (2) Use the Mann-Whitney U test to select features related to the relative efficacy of NAC treatment; (3) Rank the feature importance using a random forest model and select the top 100 most important features; (4) Use the least absolute shrinkage and selection operator (LASSO) regression and 10-fold cross-validation to select features with non-zero coefficients.

[0055] The effectiveness of the multi-sequence and multi-ROI models was verified through ablation experiments. Nine combined models were constructed by integrating the features of the selected MRI sequences or ROIs. Each model included the radiomics features of the pre-NAC and during-NAC MRI images.

[0056] Step 4. Select the best model for in-depth construction based on comprehensive performance indicators and image information: Table 2 provides the performance indicators of each radiomics model in the training set and the test set. In the training set, the DCE+ADC-tumor+peri model produced the best prediction, with an AUC of 0.864 (95% CI: 0.792 - 0.911). In the test set, the models using the DCE+ADC sequence showed relatively high and more stable performance. The DCE+ADC-tumor model had an AUC of 0.663 (95% CI: 0.549 - 0.767), the DCE+ADC-peri model had an AUC of 0.598 (95% CI: 0.502 - 0.690), and the DCE+ADC-tumor+peri model had an AUC of 0.585 (95% CI: 0.480 - 0.677). To obtain more information, the DCE+ADC-tumor+peri multi-sequence and multi-region model was selected for further radiomics model construction.

[0057] Table 2 - Performance Indicator Table for the First Construction of Radiomics

[0058]

[0059]

[0060] Step 5. In-depth construction of radiomics models (Original, Delta, Fusion): Three models were constructed using combinations of multi-sequence and multi-ROI features from different aspects of radiomics. The Original model combined the radiomics features of NAC-pre and NAC-mid. The Delta model included the Delta radiomics features from the changes from NAC-pre to NAC-mid. The Fusion model combined all the above features.

[0061] The radiomics models used the XGBoost algorithm based on the selected radiomics features. The grid search method and five-fold cross-validation were used to determine the optimal hyperparameters of the models. 4-folds (80% of the patients) were used to train the models, and the remaining (20% of the patients) were used to select the best hyperparameters.

[0062] Table 3 summarizes the AUC, accuracy, sensitivity, specificity, PPV, and NPV of each radiomics model in the training set and the test set. The AUC of the Delta model in the training set was 0.887 (95% CI: 0.816 - 0.930), and the AUC of the test set was 0.757 (95% CI: 0.683 - 0.817). Compared with the Delta model, the performance of the aggregation model was not improved, with an AUC of 0.887 (95% CI: 0.822 - 0.931) in the training set and an AUC of 0.749 (95% CI: 0.644 - 0.837) in the test set.

[0063] Table 3 - Radiomics Deepening Construction - Performance Index Table (Original, Delta, Fusion)

[0064]

[0065]

[0066] In the training set and the test set, the ROC curve of the Delta model is as Figure 3 shown. The calibration curve of the Delta model shows good consistency between the predicted and observed results of the two groups ( Figure 4 and Figure 5 ), and the Hosmer - Lemeshow goodness - of - fit test results are not significant (p = 0.343 in the training set, p = 0.464 in the test set), indicating no deviation from perfect fit. Figure 6 )

[0067] For the Delta model, decision curve analysis (DCA) shows that when the threshold probability is between 0 and 0.95 ( Figure 7 ) and between 0 and 0.71 in the test set ( Figure 8 ), the net clinical benefit is higher than 0.

[0068] Step 6: Construct a hybrid model (radiomics + clinicopathological) and a clinicopathological model: The outputs of the Original, Delta, and Fusion radiomics models for each patient were used as radiomics features. To evaluate the correlation between clinicopathological variables and the relative efficacy of NAC, we used univariate logistic regression analysis. Backward stepwise regression method based on the Akaike information criterion (AIC) was used to select and determine important variables.

[0069] A logistic regression method was used to establish a hybrid model by combining radiomics features with important clinicopathological factors. In addition, for the sake of easy comparison, a clinicopathological model containing only clinicopathological variables was also established. To select the optimal hyperparameters of the model, a grid search method combined with five-fold cross-validation was adopted. Four folds (80% of the patients) were used to train the model, and the remaining folds (20% of the patients) were used to select the best hyperparameters. The predictive ability of the model was visually demonstrated by the ROC curve, calibration curve, and decision curve.

[0070] Univariate analysis showed that the Original radiomics score (OR, 2756.20 [95% CI: 273.82–39542.24]; P < 0.001), Delta radiomics score (OR, 469.92 [95% CI: 84.13–3442.97]; P < 0.001), and Fusion radiomics score (OR, 743.78 [95% CI: 118.71–6508.12]; P < 0.001) were significantly associated with the relative efficacy of NAC in the training set, as detailed in Table 5. Multivariate analysis indicated that, in addition to the Original radiomics score, menstrual status, clinical T stage, and HER2 status were identified as independent predictors of the relative efficacy of NAC. For the Delta radiomics score, the clinical T stage was identified as an independent predictor of the relative efficacy of NAC (Table 6). Similarly, the clinical T stage also independently predicted the relative efficacy of NAC for the Fusion radiomics score.

[0071] These variables were used to construct a hybrid model to predict the relative efficacy of NAC. Table 4 provides a summary of the AUC, accuracy, sensitivity, specificity, PPV, and NPV for each hybrid model in the training set and test set. After incorporating important clinicopathological features into the radiomics model, the performance of the Delta+clinicopath model continued to be superior to that of the Original+clinicopath and Fusion+clinicopath models, with an AUC of 0.893 (95% CI: 0.877 - 0.899) in the training set and 0.780 (95% CI: 0.747 - 0.799) in the training set. Figure 3 and Figure 4 The ROC curves of the most effective hybrid model (Delta+clinicopath) and the clinicopathological model only are shown.

[0072] Table 4 - Hybrid model construction (radiomics + clinicopathology, single clinicopathology)

[0073]

[0074] Table 5 - Clinicopathological features

[0075]

[0076]

[0077] Table 6 - Delta independent predictors

[0078]

[0079] For the optimal mixed model, calibration curve analysis showed reasonable agreement between the predicted probabilities and the actual outcomes regarding the relative effectiveness of NAC in the training and test sets. The results of the Hosmer - Lemeshow test were not significant for both groups, with a p - value of 0.592 for the training set and 0.295 for the test set, indicating no significant deviation from a perfect model fit. The DCA showed that the optimal mixed model provided significant clinical net benefit at all threshold probabilities in the training set (as Figure 7 shown) and between 0 and 0.74 (in the test set (as Figure 8 shown)).

[0080] Step 7, Evaluate the comprehensive performance indicators of each model: The chi - square test was used to compare the clinicopathological characteristics of different groups or subgroups of patients with categorical variables. The performance of the model was evaluated by the area under the receiver operating characteristic curve (AUC). The sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) of the training and test sets were calculated separately. The 95% confidence interval (CI) for each metric was calculated using the bootstrap method with 1000 intervals. The optimal cut - off values for the radiation group scores in the training set were determined by maximizing the Youden index, and then these fixed cut - off values were applied to the test set. The DeLong test was used to compare the AUCs between models. All statistical tests were two - sided, and a P - value < 0.05 was considered statistically significant. All statistical analyses were performed using R 4.2.3 or Python 3.9.13.

[0081] Compared with the radiomics model, the corresponding hybrid models showed a certain degree of AUC improvement, especially the Delta+clinicopath model, with an AUC of 0.893 (95% CI: 0.877 - 0.899) in the training set and 0.780 (95% CI: 0.747 - 0.799) in the test set. In addition, both the Delta radiomics model and the Delta+clinicopath model showed good calibration and superior clinical utility in DCA and calibration curve analysis. These results confirmed that the multi-sequence, multi-region MRI radiomics model is a reliable and reproducible tool for predicting the relative efficacy of four cycles of T treatment after four cycles of EC treatment. In summary, the method for training a prediction model for predicting the relative efficacy of EC and T in NAC based on MRI radiomics and clinicopathology provided in this embodiment has the following technical effects: This embodiment provides a method for training a prediction model for predicting the relative efficacy of EC and T in NAC based on MRI radiomics and clinicopathology: a radiomics model based on DCE and ADC sequences and a hybrid model combining clinicopathological factors are constructed, and finally a Delta model integrating Original and variant radiomics features and a Delta+clinicopath hybrid model combining clinicopathological factors are obtained, which showed the best performance in predicting the relative efficacy of EC and T treatments; this method can effectively predict the relative efficacy of EC (epirubicin plus cyclophosphamide) and T (taxanes) treatment regimens in the middle stage of NAC, in order to timely adjust the treatment plan and improve the treatment effect.

[0082] It should be understood that the above-mentioned various technical means can be implemented separately in the form of hardware, software, or an integration of the two. Therefore, the methods and devices involved in the present invention, as well as several of their constituent elements or aspects, can be built into a physical medium, such as a floppy disk, CD-ROM, hard disk, or program code (i.e., a series of instructions) carried by any other machine-readable storage medium. When this program is loaded onto a device such as a computer and executed, the device is transformed into a tool for implementing the present invention.

[0083] Although the present invention has been described above with reference to the embodiments, various improvements can be made to it and components therein can be replaced with equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention can be combined with each other in any way, and the exhaustive description of these combinations is omitted in this specification only for the sake of saving space and resources. Therefore, the present invention is not limited to the specific embodiments disclosed in the text, but includes all technical solutions falling within the scope of the claims.

Claims

1. A prediction model training method for predicting the relative efficacy of EC and T in NAC, characterized in that: Here are the steps: S1. Screening clinical data: Obtain the imaging data and clinical pathological information of the subjects to construct a data set; S2. Image segmentation: The second post-contrast phase image of the DCE sequence was selected for segmentation in the six dynamic phases. The region of interest including the entire MRI-visible tumor was manually delineated layer by layer in the axial plane. Then, the intra-tumor region was three-dimensionally isotropically expanded by 5 mm to obtain the peritumoral region. S3. Feature extraction: Extract radiomic features from images. Before feature extraction, the intensity distribution of the image is normalized. Then, relevant features are extracted from each region of interest in the DCE sequence and the apparent diffusion coefficient map. The difference between the radiomic feature values ​​of the NAC-pre-stage and NAC-mid-stage images is calculated as the Delta radiomic feature to capture the longitudinal changes of the tumor. S4. Qualitative relative efficacy characteristics: The tumor volume change rate was used to characterize the relative efficacy of the initial 4 EC treatment cycles and the subsequent 4 T treatment cycles; S5. Construction of radiomics model: The model construction is divided into two parts: the first part evaluates the performance of the model using multiple sequences and multiple regions of interest through ablation experiments; the second part integrates various radiomics features to construct the Original model, Delta model and Fusion model to predict the relative efficacy of EC and T treatment using multiple sequences and multiple regions of interest; S6. Construction of hybrid model: The outputs of the original, delta, and fusion radiomics models for each patient were used as radiomics features. Logistic regression was used to combine radiomics features with important clinicopathological factors to build a hybrid model. S7. Statistical analysis and evaluation: Evaluate the performance of radiomics models and hybrid models.

2. A prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S1, the image data is MRI image data, including image data of three key points: NAC-early, NAC-mid, and NAC-late, and the image data includes at least one positioning sequence, axial TIRM sequence, DWI, fat-suppressed DCE sequence, and sagittal fat-suppressed T2-weighted imaging.

3. A prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S1, the clinical pathological data include gender, age, menstrual status, clinical T stage and clinical N stage, ER status, PR status, HER2 status and Ki67.

4. A prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S2, the region of interest containing the entire MRI-visible tumor was manually delineated layer by layer in the axial plane using 3D Slicer software.

5. The prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S2, the region within the tumor was three-dimensionally isotropically expanded by 5 mm using the SimpleITK package in Python.

6. A prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S3, PyRadiomics is used to extract radiomic features from the images.

7. A prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S4, the relative efficacy is divided into two types: the first is to determine the efficacy of 4 EC treatment cycles by calculating the relative net change of tumor volume from NAC-early stage to NAC-middle stage; the second is to determine the efficacy of 4 T treatment cycles by calculating the relative net change from NAC-middle stage to NAC-late stage, and then evaluate the relative efficacy of the treatment by comparing these net change values.

8. The prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S5, the multiple sequences and multiple regions of interest mainly use two MRI sequences: DCE and ADC and two multiple regions of interest: the intratumor region and the peritumoral region.

9. A prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S5, feature selection includes the following steps: Select features with variance greater than 1.0 through variance threshold; The Mann-Whitney U test was used to select characteristics associated with the relative efficacy of NAC treatment; The random forest model is used to rank the importance of features and select the top 100 most important features; Least absolute shrinkage and selection operator regression and 10-fold cross validation were used to select features with non-zero coefficients.

10. A prediction model training method for predicting the relative efficacy of EC and T in NAC according to claim 1, characterized in that: In step S5, the Original model combines the radiomic features of NAC-pre and NAC-metaphase, the Delta model includes the Delta radiomic features of the changes from NAC-pre to NAC-metaphase, and the Fusion model combines all of the above features.

Images