CT Radiomics for Predicting Pathological Complete Response of Axillary Lymph Nodes in Breast Cancer After Neoadjuvant Chemotherapy: A Prospective Study

Abstract

Background

The diagnostic effectiveness of traditional imaging techniques is insufficient to assess the response of lymph nodes (LNs) to neoadjuvant chemotherapy (NAC), especially for pathological complete response (pCR). A radiomics model based on computed tomography (CT) could be helpful.

Patients and Methods

Prospective consecutive breast cancer patients with positive axillary LNs initially were enrolled, who received NAC prior to surgery. Chest contrast-enhanced thin-slice CT scan was performed both before and after the NAC (recorded as the first and the second CT respectively), and on both of them, the target metastatic axillary LN was identified and demarcated layer by layer. Using pyradiomics-based software that was independently created, radiomics features were retrieved. A pairwise machine learning workflow based on Sklearn (https://scikit-learn.org/) and FeAture Explorer was created to increase diagnostic effectiveness. An effective pairwise auto encoder model was developed by the improvement of data normalization, dimensionality reduction, and features screening scheme as well as the comparison of the prediction effectiveness of the various classifiers,

Results

A total of 138 patients were enrolled, and 77 (58.7%) in the overall group achieved pCR of LN after NAC. Nine radiomics features were finally chosen for modeling. The AUCs of the training group, validation group, and test group were 0.944 (0.919-0.965), 0.962 (0.937-0.985), and 1.000 (1.000-1.000), respectively, and the corresponding accuracies were 0.891, 0.912, and 1.000.

Conclusion

The pCR of axillary LNs in breast cancer following NAC can be precisely predicted using thin-sliced enhanced chest CT-based radiomics.

Traditional imaging techniques are insufficient to assess the response of lymph nodes (LNs) to neoadjuvant chemotherapy. This article evaluates the effectiveness of a CT-based radiomics model in predicting the axillary pathological complete response after neoadjuvant chemotherapy.

Implications for Practice.

Lymph node response to NAC in breast cancer can be assessed using thin-sliced enhanced chest CT-based radiomics. The radiomics feature alterations between 2 subsequent CT images may increase the accuracy in axillary LN evaluation in actual practice. Key points include the following: (1) thin-sliced enhanced chest CT can be used to assess response of lymph nodes to neoadjunvant chemotherapy in breast cancer, (2) using auto-encoder, the changes in metastatic LNs on CT following neoadjunvant chemotherapy may be analyzed to determine whether pCR was achieved or not, and (3) the pathological treatment response might be evaluated by comparing the difference between 2 subsequent CT scans.

Introduction

The malignancy with the highest incidence in the world is breast cancer (BC).¹ Ipsilateral axillary lymph node (LN) metastases are discovered in approximately 40% of individuals with newly diagnosed breast cancer. Neoadjuvant chemotherapy (NAC) has been demonstrated to be effective in achieving pathological complete response (pCR) in 20%-70% of metastatic axillary LNs.^2-6 According to the historical meta-analysis by Cortazar et al, pCR plays a crucial predictive function, particularly in triple-negative breast cancer (TNBC).⁷ Furthermore, precise pCR determination aids in choosing the best medical strategy. As to those with LNs achieving pCR, axillary LN dissection may be avoided, which is a huge benefit for patients. For TNBC patients who did not achieve pCR after neoadjuvant therapy, the create-X clinical trial discovered extra benefits with the inclusion of adjuvant capecitabine for 6-8 cycles.⁸ Patients who did not achieve pCR had substantially longer event-free survival with the inclusion of neoadjuvant and adjuvant pembrolizumab compared to chemotherapy alone. The inclusion of post-neoadjuvant Olaparib in gBRCA mutated patients who did not achieve pCR after NAC resulted in a considerable improvement.⁹ In MonarchE clinical trial, the addition of the CDK4-6 inhibitor abemaciclib in the adjuvant setting for patients who received NAC with luminal high-risk breast cancer improved outcomes.¹⁰ In conclusion, it is crucial to precisely evaluate if the lymph nodes have achieved pCR after NAC before surgery for axillary re-staging in NAC-treated patients.

However, it has never been easy to confirm pCR without operation. Currently, ultrasonography is mostly used to evaluate axillary LNs. However, the axillary ultrasound examination is not satisfactory enough.¹¹ When compared to ultrasound, breast magnetic resonance imaging (MRI) showed a non-superior performance in terms of predicting axillary pCR post-NACT.^12,13 Additionally, sentinel lymph node biopsy (SLNB) following NAC was shown to have a false negative rate of more than 10% for patients with unselected breast cancer, according to the SENTINA and ACOSOG 1071 trials.^2,14 These approaches are all a long way from being effective.

Contrast-enhanced chest CT offers various benefits as a non-invasive imaging technique, including high temporal and spatial resolution and strong reproducibility. It is an essential examination for regular staging prior to the start of therapy. It can show off the morphology of LNs as well as how they relate to other nearby structures, including vessels in the axillary region.¹⁵ Furthermore, the blood supply of LNs could be evaluated with the use of contrast enhancement. So far, however, CT is rarely used to evaluate the response of axillary LN to NAC.

Radiomics, in conjunction with various classifiers, has the capacity to analyze both temporal and spatial heterogeneities through quantitative serial data evaluation, and it has shown promise in the treatment of breast cancer.^16-19

In this prospective study, we aimed to evaluate the effectiveness of CT-based radiomics in predicting the axillary pCR following NAC. Taking into account the contrast process in human learning, we proposed a technique for predicting pCR of axillary LNs following NAC that is based on the similarity between typical cases (template) and other cases, also known as a pairwise analysis. The prediction effectiveness of the various classifiers was also compared and evaluated.

Materials and Methods

This study was approved by our institutional review board, and patient informed consent was obtained from each patient. The study was registered (clinicaltrials.gov identifier NCT 03247478).

Patient Enrollment

Consecutive patients with breast cancer with metastatic axillary LNs were prospectively enrolled in our study. The inclusion criteria were as follows: (1) invasive breast cancer with positive axillary LNs; (2) clinical stage: T_1-3N_1-2M₀; (3) NAC was administered as scheduled; (4) mastectomy or breast-conserving surgery was performed following NAC; (5) the first chest enhanced CT was completed within 2 weeks before the start of NAC; and the second chest enhanced CT scan was within 2 weeks after the completion of NAC and within 2 weeks before the surgery. Patients with M1 stage, breast carcinoma in situ alone, synchronous bilateral breast cancer, no NAC, negative axillary LNs, missing of either the first or the second chest enhanced CT scan, poor image quality, and/or unidentification of the target LN on CT that corresponded to the biopsied node on the ultrasound were excluded.

The NAC included epirubicin+ cyclophosphamide 4 cycles (epirubicin at 100 mg/m² + cyclophosphamide at 600 mg/m² on day 1) every 2 weeks or epirubicin+ cyclophosphamide 4 cycles followed by 12 weeks of intravenous paclitaxel at 80 mg/m² on day 1, every 1 week.

All resected axillary LNs were cut and analyzed by a pathologist. No residual tumor in the final pathology for LN after NAC was considered pCR of LN; otherwise, it was considered to be non-pCR. The study complied with the ethical standards of Peking University Cancer Hospital. The study protocol was approved by our institutional review board, and each patient provided informed consent. Trastuzumab, was administered for 4 q3w cycles to HER2-positive patients in combination with paclitaxel during neoadjuvant therapy and trastuzumab was maintained alone following surgery to complete a full year of anti-HER2 treatment.

CT Scan

All contrast-enhanced chest CT scans were carried out on a 64-slice CT scanner (LightSpeed 64, GE Healthcare, Milwaukee, WI, US). The scans ranged from the apex of the lung to the level of the diaphragm. Plain scan before the administration of contrast agent and enhanced phase 25-30 s after contrast administration was performed as follows: Single breath-hold, tube current 230-445 mA, slice thickness and interval 5 mm, reconstruction thickness 0.625 mm, reconstruction interval 0, 200 kV, standard algorithm. The contrast agent (Iohexol, 300 mol mL⁻¹) was injected through the anterior cubital vein at the rate of 3-3.5 mL s⁻¹, with the total dose calculated as follows: weight (kg) × 1.2 mL kg⁻¹.

Image Processing

The software, Slicer (4.8.1, NA-MIC, NAC, Brin), was applied to review the CT images in DICOM format for each patient. Without being informed of the final pathological results of surgery, a radiologist with 15 years of expertise in breast imaging (L.Y.L.) performed all the image reviews and evaluations. The target LN was identified on the first CT based on the ultrasound’s morphological characteristics and location, matching the one that had been biopsied and was pathologically confirmed to be metastatic prior to NAC. and then the same one following NAC was identified on the second CT, taking the initial position and surrounding anatomical structure on the first CT as a reference (Fig. 1). A single target LN was chosen for each patient. The quantitative image parameters of target LN were measured and analyzed, including the long-axis diameter, short-axis diameter, area of the greatest axial section(mm²), cortical thickness(mm), CT enhancement (CT value, HU), and L/S ratio (long-axis diameter/short-axis diameter). And the target LN was delineated layer by layer and saved in the .nrrd format. Long diameter and vertical short diameter of breast tumor on axial CT images were also recorded. When an ambiguous situation occurred, a decision was made upon consultation with a senior radiologist (25 years of diagnosing experience in imaging, S.Y.S.).

Figure 1.

The color-covered region denoting the target LN, (A) CT taken before the treatment, (B) CT taken after the treatment. Long arrows on the 2 CTs are to point at the same vessels, and so was the short ones. And with the assistance of these surrounding structures, the target LN after NAC could be accurately targeted, even when it changed significantly after treatment.

Statistical Analysis

The estimated sample size was calculated as follows: 156 cases (52 cases of pCR, 104 cases of non pCR) would be required according to axillary LN pathology after dissection following NAC, pCR: non-pCR = 1:2 and case shedding 10%. (clinicaltrials.gov identifier NCT 03247478).

Positive predictive value (PPV) was defined as the number of patients with both a clinical complete response (cCR) and a pCR in axillary LNs, divided by the total number of patients with a cCR. Negative predictive value (NPV) was defined as the number of patients without a cCR and a pCR, divided by the total number of patients without a cCR.

Image characteristics screening: independent sample t-test to compare the difference in parameter distribution between pCR group and non-pCR group. Receiver operating characteristic curve (ROC) was plotted out.

Data Preprocessing and Model Training

Features Extraction

Radiomics features were retrieved using a program from an open source tool, pyradiomics (https://pyradiomics.readthedocs.io/en/latest/index.html). The categories of features extraction include first-order statistics (first-order), shape Eigenvalue (shape), gray level co-occurrence matrix (GLCM), gray level run length matrix (GLRLM), gray level size zone matrix (GLSZM), gray level difference co-occurrence matrix (GLDM), and neighborhood gray tone difference matrix (NGTDM) (Fig. 2). Image filtering techniques such as Gaussian filtering and wavelet transform were not applied due to considering the relatively small sample size. Finally, 109 features were retrieved.

Figure 2.

The process of data preprocessing and model establishment. (A) Manual delineation of the axillary LNs and 3D view; (B) Radiomics Eigenvalues extraction, including the first-order Eigenvalues, shape, and texture Eigenvalues; (C) Sample-data paring and the establishment of the model. Firstly, odd number cases with typical NAC response were selected as templates in each of the 2 patient groups, which were then paired with all samples. It was a positive sample pair if the template and the paired case were from the same group; otherwise, it was a negative sample pair. Next, we normalized the sample pair, reduced the dimension of Eigenvalues, and selected the target Eigenvalues with an optimized scheme, and finally conducted the training and testing of the AE model with sigmoid layer for classification task. Finally, the prediction label of each sample was obtained according to the probability of each sample data pair and the template label.

Model Training

To improve the prediction effectiveness of the model, several schemes in data standardization, dimension reduction, and feature screenings were compared, and the ideal amount of features in predicting the efficiency of the model was also been explored. In data standardization, 3 methods were compared and the details are below:

1.Normalize to a unit, with the following normalization formula:

$${\displaystyle \begin{array}{l}{\displaystyle {\displaystyle \overrightarrow{{\chi }_n}}=\frac{{\displaystyle \overrightarrow{x_n}}}{\sqrt{x_{1n}^2+x_{2n}^2+\cdots +x_{mn}^2}}}\end{array}}$$ (1)

($\overrightarrow{x_n}$ denotes the nth feature, m denotes the number of samples).

2. Normalize to 0-center, with the following normalization formula:

$${\displaystyle \begin{array}{l}{\displaystyle {\displaystyle \overrightarrow{x_n}}=\frac{{\displaystyle \overrightarrow{{\chi }_n}}-{\displaystyle \overline{x_n}}}{\mathrm{S}\mathrm{t}\mathrm{d}\left({\displaystyle \overrightarrow{x_n}}\right)}}\end{array}}$$ (2)

($\overrightarrow{x_n}$ denotes the mean value of the features of all samples, $\mathrm{S}\mathrm{t}\mathrm{d}\left(\overrightarrow{x_n}\right)$ denotes the standard deviation of the features).

3. Normalize to a unit with 0-center, with the following normalization formula:

$${\displaystyle \begin{array}{l}{\displaystyle {\displaystyle \overrightarrow{x_n}}=\frac{{\displaystyle \overrightarrow{x_n}}-{\displaystyle \overline{x_n}}}{\sqrt{x_{1n}^2+x_{2n}^2+\cdots +x_{mn}^2}}}\end{array}}$$ (3)

In dimensionality reduction, the effect of principal component analysis (PCA) and Pearson correlation coefficients (PCC) on prediction effectiveness were compared. The effects on the model of multivariate analysis of variance (ANOVA), recursive eigenvalue elimination (RFE), and Relief were compared at the stage of features screening schemes, and the optimal combination was eventually confirmed to establish the model.

Following preliminary tests, an auto encoder (AE) with sigmoid layer for the classification task was selected as the best classifier. The loss function is shown as

$${\displaystyle \begin{array}{l}{\displaystyle \mathrm{L}x,y=-{\displaystyle \underset{i=1}{{\displaystyle \sum ^{d_x}}}}{x}_i\log (y_i)+(1-x_i)\log (1-y_i)}\end{array}}$$ (4)

To improve the model’s accuracy and robustness of the model in a small sample set, we proposed a method based on the similarity between typical cases (template) and atypical ones for predicting the response of LNs to NAC, commonly known as contrast learning. In this analysis, we selected 7 typical pCR cases and non-pCR cases as templates and paired them with other samples to compute the differences. Variations within the same group of data were referred to as “positive cases,” whereas those between groups were referred to as “negative cases.” (Formula (5)). Finally, a new sample category was generated by voting based on the classification results of positive and negative cases as well as the label categories.

$${\displaystyle \begin{array}{l}{\displaystyle {{\displaystyle \overrightarrow{V}}}_{P(N)P}={{\displaystyle \overrightarrow{M}}}_I-{{\displaystyle \overrightarrow{N}}}_i(I\in \{1,2,3,\dots ,7\};i=1,2,3,\dots ,n)}\end{array}}$$ (5)

where ${\overrightarrow{M}}_I$and ${\overrightarrow{N}}_i$represent the Ith template and the ith sample, respectively. If the template and the sample belong to the same category, the sample pair value is positive${\overrightarrow{V}}_{PP}$. and if they belong to different categories, the sample pair value is negative ${\overrightarrow{V}}_{NP}$.

In this research, the pairwise classifier was used to describe the classifier based on paired analysis. The ideal combination of the model was selected (equations (6) and (7)) based on the accuracy, sensitivity, and specificity of the model. The analysis model was developed based on the whole procedure for creating the analytical model using Sklearn (https://scikit-learn.org/) and FeAture Explorer (FAE, v0.2.5, https://github.com/salan668/FAE) as depicted in Fig. 2.

$${\displaystyle \begin{array}{l}{\displaystyle P_i=\mathrm{A}\mathrm{v}\mathrm{g}(P_{PPi}+(1-P_{NPi}))(i=1,2,3,\dots ,n)}\end{array}}$$ (6)

$${\displaystyle \begin{array}{l}{\displaystyle N_i=\{\begin{array}{l}\mathrm{i}\mathrm{f}{P}_i<0.5\mathrm{a}\mathrm{n}\mathrm{d}{M}_I\in 1(0),\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{n}{N}_i\in 0(1)\\ \mathrm{i}\mathrm{f}{P}_i>0.5\mathrm{a}\mathrm{n}\mathrm{d}{M}_I\in 1(0),\mathrm{t}\mathrm{h}\mathrm{e}\mathrm{n}{N}_i\in 1(0)\end{array}}\end{array}}$$ (7)

The overall patient group is divided into 3 parts: training set, validation set, and test set. The total cases are chronologically divided into a training set and a test set in a 7:3 ratio, while the training set is separated into training and validation samples in a 4:1 ratio. During the training process, the 5-fold method is used for rounds of training, and the validation set is involved in training, which is equivalent to internal validation, whereas the test set has not been involved in training since the beginning, which is equivalent to external validation.

Results

Out of the 160 eligible patients, 13 were disqualified for failing to perform the second chest enhanced CT scan as scheduled following NAC, 5 for poor image quality, and 4 for failing to identify the target LN on CT that corresponded to the biopsied node on the ultrasound. Finally, 138 breast cancer patients (57 patients achieved pCR and 81 patients with non-pCR) with axillary LN metastasis were analyzed. The clinical characteristics of the patients are shown in Table 1.

Table 1.

Clinical features of the patients.

Clinical feature		Pathology after NAC		P
		pCR (n = 57)	Non-pCR (n = 81)
Age, years		48.45 ± 9.45	49.33 ± 9.11	.558
T stage	T1	10	17	.534
	T2	37	55
	T3	10	9
N stage	N1	56	77	.649
	N2	1	4
ER	Negative	23	14	.003
	Positive	34	67
PR	Negative	25	16	.002
	Positive	32	65
HER2	Negative	26	63	<.001
	Positive	31	18
Subtype	ER/PR(+), HER2(−)	15	55	<.001
	ER/PR(−), HER2(−)	11	8
	HER2(+)	31	18
Pathological grade	1	4	7	.023
	2	35	64
	3	18	10

Abbreviations: ER, estrogen receptor; HER2, human epidermal growth factor receptor-2; LN, lymph node; PR, progesterone receptor.

Finally, 102 patients (42 pCR patients and 60 non-pCR patients) were divided into the training set and test set, and 36 patients (15 pCR patients and 21 non-pCR patients) were divided into the test set. All characteristics were comparable between the training set and the test set.

Correlation of CT Parameters and Pathological Results After NAC

The diagnostic effectiveness of the target LNs’ CT parameters on the first CT taken before NAC and the second CT after NAC for predicting pCR of LNs are shown in Table 2.

Table 2.

Comparisons of CT parameters between different pathological remission status of target LNs before and after NAC.

	Image parameters	Pathology after NAC		P value	AUC
		pCR	Non-pCR
The first CT	LN long-axis diameter (mm)	18.99 ± 7.08	19.51 ± 10.20	.508	0.531
	LN short-axis diameter (mm)	12.61 ± 5.45	13.49 ± 6.42	.624	0.529
	LN L/S ratio	1.58 ± 0.44	1.46 ± 0.33	.104	0.582
	Tumor long-axis diameter (mm)	24.76 ± 10.91	26.25 ± 13.06	.485	0.509
	Tumor short-axis diameter (mm)	17.61 ± 6.24	17.68 ± 7.17	.955	0.521
	Area (mm2)	270.12 ± 217.58	321.11 ± 367.09	.588	0.604
	CT enhancement (HU)	69.13 ± 19.01	68.07 ± 19.33	.955	0.510
	Cortical thickness (mm)	11.88 ± 6.07	13.79 ± 13.29	.517	0.534
The second CT	LN long-axis diameter (mm)	9.43 ± 4.31	11.54 ± 7.18	.103	0.573
	LN short-axis diameter (mm)	5.44 ± 2.29	7.18 ± 4.62	.007	0.643
	L/S ratio	1.90 ± 0.95	1.67 ± 0.56	.168	0.569
	Tumor long-axis diameter (mm)	11.02 ± 8.52	15.17 ± 7.72	.004	0.695
	Tumor short-axis diameter (mm)	7.11 ± 4.14	9.64 ± 4.36	.001	0.701
	Area (mm2)	57.85 ± 47.18	111.88 ± 218.31	.025	0.625
	CT enhancement (HU)	50.45 ± 23.03	58.36 ± 20.27	.036	0.598
	Cortical thickness (mm)	4.29 ± 2.07	6.23 ± 4.79	.001	0.670
Changes between the second CT and the first CT	LN long-axis diameter (mm)	−0.47 ± 0.25	-0.39 ± 0.20	.005	0.624
	LN short-axis diameter (mm)	−0.53 ± 0.20	-0.44 ± 0.20	.007	0.628
	L/S ratio	0.22 ± 0.46	0.17 ± 0.40	.676	0.521
	Tumor long-axis diameter (mm)				0.679
	Tumor short-axis diameter (mm)				0.711
	Area (mm2)	−0.72 ± 0.26	-0.63 ± 0.21	.004	0.626
	CT enhancement (HU)	−0.24 ± 0.35	-0.07 ± 0.50	.022	0.661
	Cortical thickness (mm)	−0.58 ± 0.18	-0.48 ± 0.21	.030	0.627

Bold italic indicates P < .05.

Abbreviations: HU, hounsfield unit; L/S, long-axis diameter/short-axis diameter.

Significant variations existed between pCR and non-pCR patients in the second CT parameters. Patients with pCR showed more substantial differences between the 2 CTs than those with non-pCR. However, the diagnostic performance of these CT parameters for predicting pCR was subpar, with AUC ranging from 0.598 to 0.711.

Diagnostic Efficiency of Pair-Wise Auto Encoder

Normalize to 0-enter, PCC, and RFE to pre-process the features chosen in the model established after comparing and analyzing combinations of data normalization techniques, feature dimensionality reduction strategies, and feature screening schemes. The model’s prediction accuracy was at its maximum with 9 features (Fig. 3, Supplementary Fig. S1). The AUCs of the training group, validation group, and test group were 0.944 (0.919-0.965), 0.962 (0.937-0.985), and 1.000 (1.000-1.000), respectively, and the corresponding accuracies were 0.891, 0.912, and 1.000. The PPV and NPV were 0.978 and 0.834 (training group), 0.853 and 1.000 (validation group), and 1.000 and 1.000 (test group), respectively (Figs. 4 and 5).

Figure 3.

The model’ prediction AUC, the accuracy was at its maximum with 9 features.

Figure 4.

The selected Eigenvalues and their contributions in the model after normalization and dimension reduction. CV train: Result under the training data via a 5-fold cross-validation method; CV validation: Result under validation data via a 5-fold cross-validation method; Train: Result using all training data; Test: result using all test data (CV, cross-validation).

Figure 5.

The AUC values of ROC on CV training, CV validation, training, and testing data. CV train: Result under the training data via a 5-fold cross-validation method; CV validation: Result under validation data via a 5-fold cross-validation method; Train: Result using all training data; Test: result using all test data.

Discussion

Radiomics has demonstrated great value in differential diagnosis, treatment response evaluation, and prognosis prediction. Our findings reveal that as compared with the conventional quantitative CT characteristics, pairwise AE based on radiomics of contrast-enhancement CT images before NAC has a substantially greater diagnostic effectiveness. The AUC for the training group with AE was 0.981, whereas it was 0.971 for the verification group. The diagnostic efficiency was suitable for clinical use. To our knowledge, this is the first research to apply CT radiomics to evaluate the response of metastatic lymph nodes.

The prediction of the axillary pCR based on the characteristics of a single CT scan before chemotherapy was insufficient.²⁰ This might be attributable to individual interfering factors, such as varied chemotherapy regimens and unique individual responses, etc. However, when changes between 2 consecutive examinations are taken to evaluate the response, the influence of the interfering factors is reduced to some extent and the trend may be tracked based on the changes.^21,22 Furthermore, with the use of contrast enhancement, thin-layer technology, and image reconstruction, CT can well display the detailed structures of lymph nodes, even in the case of relatively small nodes with evident shrinkage and distortion following treatment. We found that the enhancement of LNs showed a similar diagnostic efficiency to that of the diameters (AUC 0.627), indicating that when the treatment is effective, the blood supply to LNs would likewise be greatly reduced. Our findings demonstrated that substantial differences existed between the pCR and non-pCR of the metastatic LN group concerning the changes in the long-axis diameter, the short-axis diameter, and the cortical thickness between 2 consecutive CT scans. The short-axis diameter, area, and enhanced CT strength in preoperative CT (the second CT) following NAC, also showed significant statistical differences, indicating that the preoperative CT characteristics following NAC have a direct relationship with treatment response. This might be owing to the short-time interval between preoperative CT and surgery, as well as the lack of extra therapy influence. However, the sensitivity and diagnostic accuracy of either a single CT or changes in parameters between 2 CTs were still insufficient. The results showed that the same measurable parameters on CT as on the ultrasound (the long-axis diameter, the short-axis diameter, area, and cortical thickness) had comparable and acceptable performances in evaluating the treatment response with an AUC of >0.6.^22–24 The short-axis diameter and the long-axis diameter of metastatic LNs performed similarly in predicting pCR of LNs (AUC 0.628 vs. 0.624), while cortical thickness performed the best performance among all CT parameters (AUC 0.661).

Recently, several studies have used breast MRI-based radiomics to predict axillary LN metastases in patients undergoing direct surgery.^25,26 However, the studies concentrated on the indirect characteristics of breast lesions in evaluating axillary LNs rather than the characteristics of the LNs themselves when analyzing LNs.

Our research specifically targeted metastatic axillary lymph nodes, and we found that combing the radiomics characteristics of CTs taken before and after NAC treatment significantly improved the diagnostic effect. An AE model, which has been proven to be a good algorithm in data processing,^27,28 was selected after the prediction performance comparison of varied classifiers. Meanwhile, considering the contrast process in human learning, we suggested a method for predicting the treatment response directly via analyzing the metastatic LN itself that is based on the similarity between typical cases (template) and other cases and also known as pairwise learning. This measure was designed to make full use of the correlation information between the 2 types of samples and within each sample, and also to enhance the samples themselves. The greater test set performance in our study indicates that the experiment might have a reduced degree of overfitting and stronger generalization capacity even with limited samples.

After weighing the effects of varied data preprocessing techniques on the classification effect throughout the modeling process, we finally settled on normalize to 0-center, PCC, and RFE to carry out data standardization, feature reduction, and feature screening. The Relief algorithm established a relationship between features and categories based on the features’ discernabilitiy in near samples. Higher test set performance results indicate that the experiment has less over-fitting and better generalization capacity. Several researches have studied the diagnostic performance of radiomics of positron emission tomography (PET)/CT in predicting axillary pCR.^29,30 Due to the full use of relevant information between cases and the augmentation of samples in pairwise analysis, the paired AE model based on Contrast-enhanced CT radiomics before and after NAC demonstrated equal prediction performance in our research.

Nevertheless, there are several limitations in this research. First of all, the sample size was relatively limited and there were not enough cases of each immunohistochemical subtype to carry out statistical analysis based on each immunohistochemical subtype. Second, there are many methods for radiomics analysis, and the results obtained by different methods might vary, which needs to be further compared and discussed. Third, intelligent image analysis needs to be performed based on manual image segmentation currently, which is time-consuming. Automated intelligent image recognition and segmentation are anticipated to be possible with technological improvements, which might reduce labor requirements and save time.

Conclusion

The model based on CT-based radiomics before and after NAC may significantly improve the evaluation efficiency and provide considerable auxiliary potential for reliably identifying patients that achieved axillary LN pCR.

Funding

This work was supported by the Science Foundation of Peking University Cancer Hospital (2020-21), Beijing excellent talents training project (2018000021469G260), PKU-Baidu Fund (No. 2020BD027), Beijing Municipal Science and Technology Project(Z181100001918001), Beijing Municipal Administration of Hospitals Clinical Medicine Development of Special Funding Support (No. ZYLX201803), and Beijing Hospitals Authority Ascent Plan (DFL20191103).

Conflict of Interest

The authors indicated no financial relationships.

Author Contributions

Conception/design: Y.-L.L., L.-Z.W., T.O., Y.-S.S. Development of methodology: Q.-L.S., H.-T.Z. Collection and/or assembly of data: Y.-L.L., J.-F.L., T.-F.W., Y.-S.S. Administrative, technical, or material support (reporting or organizing data, constructing databases): T.O., Y.-S.S. Data analysis and interpretation: Q.-L.S., Y.-J.H., X.-T.L., Z.-Q.F. Writing, review, and/or revision of the manuscript: Y.-L.L., L.-Z.W., T.O., Y.-S.S. Final approval of manuscript: All authors.

Data Availability

Data can be acquired upon reasonable request to corresponding author by email. The code part is an open source and can be uploaded to Geithub for interested researchers to download.