Rationale and Objectives
To predict mutations in TP53 and PIK3CA genes in breast cancer using ultrasound (US) signatures and clinicopathology.
Materials and Methods
In this study, we developed and trained a model in 386 breast cancer patients to predict TP53 and PIK3CA mutations. The clinicopathological and US characteristics (including two-dimensional and color Doppler US) were investigated. Statistically significant variables were used to build predictive models, then a combined model was developed using the multivariate logistic regression analysis.
Results
Univariate and multivariate analyses revealed that calcifications on US was an independent predictor of TP53 mutation (p < 0.05), whereas diameter on US and US type were independent predictors of PIK3CA mutation in breast cancer (all p < 0.05). Meanwhile, Luminal B/Human epidermal growth factor receptor two-positive (HER2+), HER2+/estrogen receptor-negative (ER–), and triple-negative breast cancer (TNBC) subtypes were strong predictors of TP53 mutation (odds ratio [OR] = 3.13, 3.18, 3.44, respectively, all p < 0.05). HER2+/ER– and TNBC subtypes were negative predictors of PIK3CA mutation (OR = 0.223, 0.241, respectively, all p < 0.05). The areas under curves (AUCs) for PIK3CA mutation in the training set increased from 0.553-0.610 to 0.741 in the multivariate model that combined US features and molecular subtype, with a sensitivity and specificity of 80.6% and 58.7%, respectively. The application of the multivariate model in the validation set achieved acceptable discrimination (AUC = 0.715). For TP53 mutation, the AUC was 0.653.
Conclusion
US is a non-invasive modality to recognize the presence of TP53 and PIK3CA mutation. The models combined with US features and molecular subtype have implications for the practical application of predicting gene mutation for individual decision-making regarding treatment planning.
Key Words
Abbreviations:
US (ultrasound), AUCs (areas under curves), HER2 (human epidermal growth factor 2), ER (estrogen receptor), PR (progesterone receptor), ROC (receiver operator characteristic)INTRODUCTION
The TP53 and PIK3CA genes are mutated in 20-40% of human breast cancers (
1
). The TP53 regulates the expression of the P53 tumor suppressor in response to various stress signals and suppression of cellular transcriptional activity by modulating cell cycle arrest and apoptosis. The PIK3CA, which encodes the p110α catalytic subunit of phosphoinositide-3 kinases (PI3Ks), could regulate pathway activity to drive oncogenic transformation, including proliferation, apoptosis, and metabolism (2
). Recent genomic analyses have revealed that these mutations in cancer cause adverse clinical outcomes, including resistance to therapy and decreased overall survival (3
). With the introduction of genomic profiling in precision medicine, cancer treatment decisions are increasingly based not only on the patient's clinical characteristics and tumor pathology but also on individual mutational profiles (4
,5
). For example, the use of alpelisib that target specific mutation within the PIK3CA gene has improved outcomes in patients with endocrine resistance Hormone receptor-positive (HR+)/human epidermal growth factor two-negative (HER2–) breast cancer patient in which the mutation occurs (6
). Consequently, it is necessary to evaluate the gene mutation to identify patients who will benefit from the precision treatment (- De Mattos-Arruda L
PIK3CA mutation inhibition in hormone receptor-positive breast cancer: time has come.
ESMO Open. 2020; 5https://doi.org/10.1136/esmoopen-2020-000890
7
,8
).Although significant advances in sequencing technology biopsies can be informative and have become standard of care in clinical and research settings (
9
), sequencing technology still has a long turnaround time, is relatively high cost, and fails to determine mutational status due to poor DNA quality. Taken together with the dynamic changes of genetic codes in breast cancer, there is a strong demand for accurate and non-invasive means in differentiating gene mutations, which presents a unique opportunity for medical imaging.Since medical images can effectively reflect the underlying pathophysiologic characteristics, quantitative analyses are imperative to understand the biology of the digital images (
10
). Recent radiologic studies have demonstrated the relevance between radiologic information and somatic mutations in lung cancer (11
), glioblastoma (12
), as well as head and neck cancers (13
). Although imaging features were found to be significantly associated with mutational data in breast cancer, most studies either employed a small cohort size or did not include external validation, which may have compromised the observed outcomes (14
, 15
, - Li H
- Giger ML
- Huynh BQ
- et al.
Deep learning in breast cancer risk assessment: evaluation of convolutional neural networks on a clinical dataset of full-field digital mammograms.
J Med Imaging (Bellingham). 2017; 4041304https://doi.org/10.1117/1.JMI.4.4.041304
16
). However, only a handful of studies have reported using radiogenomics to analyze the association between ultrasound (US) phenotypes and somatic mutation in breast cancer.US is an excellent tool for breast cancer imaging, owing to the non-invasive nature of its procedures and the efficient characterization of tissue properties (
17
,18
). When compared with mammography and magnetic resonance imaging (MRI), it employs non-ionizing radiation and is not limited to patients with claustrophobia, pacemakers metallic objects (19
). Moreover, US has shown potential in integrating both radio-phenotypes and genotypes (20
,21
).The mutation spectrum is subtype-specific in breast cancer, suggesting the involvement of individual mutations in the pathogenesis of different breast cancer subtypes. For example, luminal-like breast cancer shows higher prevalence of PIK3CA, MAP3K1, GATA3 mutations (
22
, 23
, 24
), while HER2+/estrogen receptor-negative (ER–) and triple-negative breast cancer (TNBC) tumors have a high frequency of TP53 mutation (24
). Additionally, TNBC patients carry a high rate of BRCA mutation, especially in BRCA1 (25
).In the present study, we investigated the relationship between US features and somatic mutations in breast cancer. We developed and validated a prediction model based on US features and clinicopathological characteristics.
MATERIAL AND METHODS
Study Participants
This prospective study was approved by the Ethical Committee of the Shanghai Cancer Center of Fudan University (FUSCC). We enrolled consecutive patients who were scheduled to undergo US-guided biopsy for suspected breast cancer. Those who were recruited between January and August 2019 were included in the training set, while those who were recruited between September and December 2019 comprised the validation set. Written informed consent was obtained from all patients prior to inclusion in the study, with all patients undergoing breast US imaging prior to biopsy. Treatment plans were established by a multidisciplinary team of breast cancer specialists. Thirty-two of the 576 patients were excluded due to ipsilateral breast excision, while 23 were excluded for image indistinction. Among the 401 participants with malignant breast mass, 386 (96.3%) patients who consented to gene sequencing were included in our study and were subsequently subjected to the examination of US imaging features and genetic alteration (Fig 1).
Figure 1Flowchart of the procedure of patient recruitment, exclusions, and numbers for final grouping.
Histopathological analysis
The biopsies were routinely processed to evaluate the histological and biological characteristics of breast cancer. ER and progesterone receptor, as well as the Ki-67 index, were expressed as the percentage of cells that stained positively on immunohistochemistry (IHC). HER2 positivity was defined as 3+ using IHC and/or fluorescence in situ hybridization. Surrogate molecular subtypes were determined based on the 2011 St. Gallen guidelines (
26
).US Imaging and Analysis
A standardized breast US examination, including B-mode and vascular US examinations, was performed by a certified radiologist (J.Z.) who had 8 years of experience in breast imaging. Briefly, we used a US system (Aixplorer, Supersonic Imagine, Aix-en-Provence, France) equipped with a 4-15 MHz linear array transducer. The imaging acquisition standards were as follows: 12 conventional US images were captured starting with the largest cross-section of the tumor at equal intervals in a 180° clockwise range (Fig 2).
Figure 2Illustration shows overview of the development and validation of precidtion models by ultrasound and clinicopathological features. The US imaging were captured starting with largest cross-section of the tumor at every 15° intervals in a 180° clockwise range. ROC, receiver operator characteristic; US, ultrasound.
The two radiologists (Y.X.H and J.L.) analyzed the US images together according to the Breast Imaging Reporting and Data System of the American College of Radiology. For any disagreement, another radiologist (S.C.Z) evaluated the image until the three radiologists reached a consensus (
17
). B-mode phenotypes were dichotomized according to size, ultrasound type (mass, non-mass, or mass complicated by non-mass), shape (irregular, oval, or round), echo pattern (complex cystic and solid, hypoechoic, hyperechoic, or heterogeneous), margin (distinct, or indistinct), posterior features (no-change, enhanced, attenuation, or enhanced complicated by attenuation), and calcifications (present or absent). In addition, microvascular US phenotypes included vessel morphologic feature (complex [branching or shunting], none, or simple [dot-like or linear]), distribution (central [any vessel detected within the lesion], none, or peripheral [all vessels located at the margin]), penetrating vessels (present or absent), and perfusion detection (present or absent).Analysis of Mutations in TP53 and PIK3CA Genes
Briefly, biopsy samples were subjected to next-generation sequencing panel as described in our previous study (
27
). Somatic mutations, including frame shift deletions and insertions, in-frame deletions, insertions, as well as missense, nonsense, and splice site modifications in the TP53 and PIK3CA genes, were analysed.Statistical Analysis
All statistical analyses were performed using packages implemented in R software (V.3.6.2, http://www.r-project.org) and SPSS 23.0 (SPSS Inc., Chicago, IL, USA). Univariate analysis was performed using Student's t-test, Pearson's χ2 or Fisher's exact test, as appropriate. The features with p-values of <0.1 after univariate analysis were used to perform multivariate logistic regression with the overarching goal of selecting the most valuable variables (p < 0.05) via stepwise backward selection. The areas under curves (AUCs) of the receiver operator characteristic (ROC) were calculated to evaluate model performance and prognostic value. Next, the AUCs were compared using the Delong test, and the optimal cut-offs were obtained by maximizing the Youden index (sensitivity + specificity – 1). Next, we applied the Hosmer-Lemeshow goodness of fit test to evaluate the overall fit of the logistic model, then generated ROC curves to assess the diagnostic efficiency of the models. After that, the heat maps were generated using the hierarchical clustering algorithm to reveal the distribution ultrasound features of TP53 or PIK3CA mutation.
Inter-observer agreement for the two radiologists (Y.X.H. and J.L.) on the evaluation of US categorical and continuous variables were assessed by Cohen's κ statistics and intra-class correlation coefficient (ICC), respectively. The results of ≥0.81, 0.61-0.80, 0.41-0.60, 0.21-0.40, and ≤0.20 were considered as excellent, good, moderate, fair, and poor, respectively (
28
).RESULTS
Participants’ Clinical and Pathological Characteristics
Our study included 386 patients ranging from 26 to 75 years of age who were pathologically confirmed breast cancer at FUSCC. The patients’ clinicopathological parameters in the training (n = 243) and validation (n = 143) sets are listed in Table 1 and Table S1. In the training set, 110 (45.3%) patients had TP53 mutation, and 93 (38.3%) had PIK3CA mutation. The obtained results showed that molecular subtype differed significantly between TP53 mutant and wild-type groups in the training set (p = 0.006), and that differed significantly between PIK3CA mutant and wild-type groups in the training set (p = 0.003). TP53 exhibited high mutation rates in the HER2+/ER– (55.1%), TNBC (57.4%), and Luminal B/HER2+ (52.9%) subtypes. On the other hand, PIK3CA exhibited high mutation rates in the Luminal A (48.9%) and Luminal B/HER2– (52.5%) subtypes (Fig 3). There were no notable differences in the remaining clinicopathological features between the mutant and wild-type groups (all p > 0.1).
Table 1Clinicopathological Characteristics in the Training Set
| Characteristics | Training Set (n = 243) | ||||||
|---|---|---|---|---|---|---|---|
| TP53 | p | PIK3CA | p | ||||
| Mutant (n = 110) | Wild (n = 133) | Mutant (n = 93) | Wild (n = 150) | ||||
| Age (year) | <40/>70 | 48 (43.6) | 57 (2.9) | 0.903 | 42 (45.2) | 63 (42.0) | 0.629 |
| ≤40≤70 | 62 (56.4) | 76 (57.1) | 51 (54.8) | 87 (58.0) | |||
| Menopausal status (%) | Premenopause | 35 (31.8) | 43 (32.3) | 0.932 | 27 (29.0) | 51 (34.0) | 0.420 |
| Postmenopause | 75 (68.2) | 90 (67.7) | 66 (71.0) | 99 (66.0) | |||
| Molecular subtype (%) | Luminal A | 12 (10.9) | 33 (24.8) | 0.006* | 22 (23.7) | 23 (15.3) | 0.003 |
| Luminal B/HER2- | 22 (20.0) | 39 (29.3) | 32 (34.4) | 29 (19.3) | |||
| Luminal B/HER2+ | 18 (16.4) | 16 (12.0) | 14 (15.1) | 20 (13.3) | |||
| HER2+/ER- | 27 (24.5) | 22 (16.5) | 12 (12.9) | 37 (24.7) | |||
| TN | 31 (28.2) | 23 (17.3) | 13 (14.0) | 41 (27.3) | |||
| Histology (%) | Ductal | 98 (89.1) | 117 (88.0) | 0.679 | 77 (82.8) | 138 (92.0) | 0.139 |
| Lobular | 2 (1.8) | 6 (4.5) | 4 (4.3) | 4 (2.7) | |||
| Mix | 4 (3.6) | 4 (3.0) | 5 (5.4) | 3 (2.0) | |||
| Other | 6 (5.5) | 6 (4.5) | 7 (7.5) | 5 (3.3) | |||
| (p)N stage (%) | pN0 | 76 (69.1) | 89 (66.9) | 0.639 | 65 (69.9) | 100 (66.7) | 0.558 |
| pN1 | 26 (23.6) | 28 (21.1) | 22 (23.7) | 32 (21.3) | |||
| pN2 | 6 (5.5) | 11 (8.3) | 4 (4.3) | 13 (8.7) | |||
| pN3 | 2 (1.8) | 5 (3.8) | 2 (2.2) | 5 (3.3) | |||
| Pathology grade (%) | Unknown | 10 (9.1) | 6 (4.5) | 0.292 | 10 (10.8) | 6 (4.0) | 0.189 |
| I | 4 (3.6) | 7 (5.3) | 3 (3.2) | 8 (5.3) | |||
| II | 45 (40.9) | 66 (49.6) | 40 (43.0) | 71 (47.3) | |||
| III | 51 (46.4) | 54 (40.6) | 40 (43.0) | 65 (43.3) | |||
ER, estrogen receptor; HER2, human epidermal growth factor receptor-2; TN, triple negative.
Data are number of patients. Data in parentheses are percentages. p values for difference were determined by chi-square test or fisher's exact test for categorical variables.
Figure 3Graph showing ultrasound and molecular subtype characteristics of PIK3CA mutation and wild-type group in the training (n = 243) and validation (n = 143) sets. Status of mutation, molecular subtype, and ultrasound type are provided as labeled. Diameter for each patient is listed as continuous variable from low to high in the respective data sets.
In the validation set, 59 (41.3%) patients had TP53 mutation, and 48 (33.6%) patients had PIK3CA mutation. The results showed that molecular subtype differed significantly between TP53 mutant and wild-type groups (p = 0.013), and that differed significantly between PIK3CA mutant and wild-type groups (p = 0.008). TP53 exhibited high mutation rates in TNBC (66.7%), HER2+/ER– (60.9%), and Luminal B/HER2+ (60.9%) subtypes. On the other hand, PIK3CA exhibited high mutation rates in the Luminal A (52.4%), Luminal B/HER2– (47.6%) subtypes (Fig 3). There were no notable differences in the remaining clinicopathological features between the mutant and wild-type groups (all p > 0.1).
Observer Agreement of US Examination
The Cohen's κ statistic revealed excellent agreements in categorical variables and the levels ranged between 0.801 and 0.965 (Table S2). Meanwhile, ICC of 0.898 revealed good agreement in the continuous variable (Table S3).
Sonographic Feature Characteristics
The US characteristics between the mutant and wild-type groups in the training and validation sets are listed in Table 2 and Table S4. The obtained results indicated that significant differences (p < 0.1) were found in calcifications on US between the TP53 mutant and wild-type groups. In addition, diameter on US, US type, and US shape were significantly different between the PIK3CA mutant and the wild-type groups. In contrast, there were no visible differences of the remaining US features between the mutant and wild-type groups (all p > 0.1).
Table 2Comparison of the US Features Between Mutant and Wild-Type Groups in the Training Set
| Characteristics | Training Set (n = 243) | ||||||
|---|---|---|---|---|---|---|---|
| TP53 | p | PIK3CA | p | ||||
| Mutant (n = 110) | Wild (n = 133) | Mutant (n = 93) | Wild (n = 150) | ||||
| Diameter (mean ± sd, mm) | 21.25±9.3733 | 21.52±9.891 | 0.832 | 23.38±9.527 | 20.17±9.538 | 0.012 | |
| US type (%) | Mass | 97 (88.3) | 117 (88.0) | 0.221 | 86 (92.5) | 128 (85.3) | 0.048 |
| Non-mass | 9 (8.2) | 15 (11.3) | 4 (4.3) | 20 (13.3) | |||
| Mass + non-mass | 4 (3.6) | 1 (0.8) | 3 (3.2) | 2 (1.3) | |||
| Shape (%) | Regular | 4 (3.6) | 5 (3.8) | 0.960 | 1 (1.1) | 8 (5.3) | 0.088 |
| Unregular | 106 (96.4) | 128 (96.2) | 92 (98.9) | 142 (94.7) | |||
| Margin (%) | Distinct | 28 (25.5) | 37 (27.8) | 0.678 | 24 (25.8) | 41 (27.3) | 0.794 |
| Indistinct | 82 (74.5) | 96 (72.2) | 69 (74.2) | 109 (72.7) | |||
| Echogenicity (%) | Complex* | 11 (10.0) | 11 (8.3) | 0.388 | 10 (10.8) | 12 (8.0) | 0.802 |
| Hypoechoic | 39 (35.5) | 61 (45.9) | 40 (43.0) | 60 (40.0) | |||
| Hyperechoic | 1 (0.9) | 2 (1.5) | 1 (1.1) | 2 (1.3) | |||
| Heterogeneous | 59 (53.6) | 59 (44.4) | 42 (45.2) | 76 (50.7) | |||
| Posterior features (%) | Nochange | 42 (38.2) | 62 (46.6) | 0.123 | 40 (43.0) | 64 (42.7) | 0.760 |
| Enhanced | 31 (28.2) | 24 (18.0) | 18 (19.4) | 37 (24.7) | |||
| Attenuation | 24 (21.8) | 37 (27.8) | 25 (26.9) | 36 (24.0) | |||
| Enhanced+ attenuation | 13 (11.8) | 10 (7.5) | 10 (10.8) | 13 (8.7) | |||
| Calcification (%) | Absent | 56 (50.9) | 87 (65.4) | 0.022 | 56 (60.2) | 87 (58.0) | 0.733 |
| Present | 54 (49.1) | 46 (34.6) | 37 (39.8) | 63 (42.0) | |||
| Morphologic feature (%) | Complex | 55 (50.0) | 66 (49.6) | 0.931 | 51 (49.5) | 70 (46.7) | 0.724 |
| None | 10 (9.1) | 14 (10.5) | 8 (7.8) | 16 (10.7) | |||
| Simple | 45 (40.9) | 53 (39.8) | 44 (42.7) | 64 (42.7) | |||
| Distribution (%) | Central | 24 (21.8) | 30 (22.6) | 0.190 | 23 (24.7) | 31 (20.7) | 0.117 |
| None | 14 (12.7) | 8 (6.0) | 4 (4.3) | 18 (12.0) | |||
| Peripheral | 72 (65.5) | 95 (71.4) | 66 (71.0) | 101 (67.3) | |||
| Penetrating vessels (%) | Present | 42 (38.2) | 51 (38.3) | 0.979 | 33 (35.5) | 60 (40.0) | 0.481 |
| Absent | 68 (61.8) | 82 (61.7) | 60 (64.5) | 90 (60.0) | |||
US, ultrasound.
Except for diameter (mean), data are number of patients. Data in parentheses are percentages or standard deviations. p values for difference were determined by chi-square test or fisher's exact test for categorical variables.
Performance of the Sonographic and Clinicopathologic Features in Differentiating TP53/PIK3CA Mutations
Results obtained from univariate logistic regression analysis indicated that the molecular subtype and calcifications were significant predictors for TP53 mutation. The molecular subtype, diameter, and US type were significant predictors for PIK3CA mutation (Table 3). Then multivariate logistic regression models were generated as model 1, including molecular subtype and calcifications on US, and model 2, including molecular subtype, diameter, and US type. The predictive effect of every single feature on identifying TP53/PIK3CA status was also evaluated, which was compared with the multivariate regression models.
Table 3Results From Univariate and Multivariate Logistic Regression Analysis of Factors Used to Predict TP53 and PIK3CA Mutations
| TP53 Mutation | PIK3CA Mutation | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|
| Variables | Univariate Analysis OR (95% CI) | p | Multivariate Analysis OR (95% CI) | p | Univariate Analysis OR (95% CI) | p | Multivariate Analysis OR (95% CI) | p | ||
| Subtype | Luminal A | 1 | 1 | 1 | 1 | |||||
| Luminal B/HER2- | 1.55 (0.67-3.60) | 0.307 | 1.45 (0.62-3.39) | 0.395 | 1.15 (0.53-2.49) | 0.716 | 1.245 (0.550-2.841) | 0.600 | ||
| Luminal B/HER+ | 3.09 (1.20-7.95) | 0.019 | 3.13 (1.21-8.11) | 0.019 | 0.73 (0.30-1.80) | 0.469 | 0.668 (0.260-1.703) | 0.400 | ||
| HER2+/ER- | 3.37 (1.42-8.04) | 0.006 | 3.18 (1.33-7.65) | 0.010 | 0.34 (0.14-0.81) | 0.015 | 0.233 (0.09-0.590) | 0.003 | ||
| TN | 3.71 (1.58-8.70) | 0.003 | 3.44 (1.45-8.14) | 0.005 | 0.33 (0.14-0.78) | 0.011 | 0.241 (0.09-0.590) | 0.002 | ||
| Calcification | No | 1 | 1 | - | - | - | ||||
| Yes | 1.82 (1.09-3.06) | 0.023 | 1.76 (1.03-3.01) | 0.038 | - | - | - | |||
| Diameter | - | - | - | - | 1.04 (1.01-1.06) | 0.013 | 1.052 (1.021-1.085) | 0.001 | ||
| US type | Mass | - | - | - | - | 1 | 1 | - | ||
| Non-mass | - | - | - | - | 0.3 (0.10-0.90) | 0.032 | 0.159 (0.0412-0.480) | 0.003 | ||
| Mass + non-mass | - | - | - | - | 2.23 (0.37-13.64) | 0.384 | 3.405 (0.530-27.326) | 0.200 | ||
| Shape | Regular | - | - | - | - | 1 | - | - | ||
| Unregular | - | - | - | - | 5.18 (0.64-42.13) | 0.124 | - | - | ||
CI, confidence interval; ER, estrogen receptor; HER2, human epidermal growth factor receptor-2; OR, odds ratio; TN, triple negative; US, ultrasound.
In the prediction of TP53 mutation, the performances of calcification on US (AUC = 0.573, 95% confidence interval [CI] = [0.511-0.63]) and molecular subtype (AUC = 0.631, 95% CI = [0.563-0.700]) were slightly improved by the combination of them (model 1, AUC = 0.653, 95% CI = [0.585-0.722], Table 4) in the training set. Additionally, the difference was significant between calcification on US and model1 (DeLong test, p = 0.021), while the difference was not significant between molecular subtype and model1 (DeLong test, p = 0.171, Fig 4a). Moreover, no improvement in performance was found by a combination of the calcification on US and molecular subtypes (DeLong test, p = 0.384, p = 0.075, respectively, Fig 4b).
Table 4Predictive Performance of Models in the Training and Validation Sets
| Model | Training Set | Validation Set | |||||||
|---|---|---|---|---|---|---|---|---|---|
| AUC (95% CI) | Sensitivity (%) | Specificity (%) | AUC (95% CI) | Sensitivity (%) | Specificity (%) | ||||
| TP53 | US calcification | 0.573 (0.511-0.635) | 49.1 | 65.4 | 0.480 (0.397-0.563) | 30.5 | 42.9 | ||
| Molecular subtype | 0.631 (0.563-0.700) | 69.1 | 54.1 | 0.626 (0.535-0.717) | 42.4 | 77.4 | |||
| Model 1 | 0.653 (0.585-0.722) | 81.0 | 43.6 | 0.590 (0.496-0.685) | 27.1 | 88.1 | |||
| PIK3CA | US diameter | 0.610 (0.533-0.681) | 51.6 | 70.0 | 0.627 (0.529-0.725) | 58.3 | 66.3 | ||
| US type | 0.553 (0.514-0.591) | 95.7 | 13.3 | 0.670 (0.568-0.758) | 64.6 | 63.3 | |||
| Molecular subtype | 0.642 (0.572-0.712) | 73.1 | 52.0 | 0.536 (0.485-0.587) | 93.8 | 11.6 | |||
| Model 2 | 0.741 (0.678-0.804) | 80.6 | 58.7 | 0.715 (0.626-0.805) | 80.6 | 62.1 | |||
AUC, areas under curves; CI, confidence interval; US, ultrasound.
Model 1, combination of US calcification and molecular subtype; Model 2 combination of US diameter, US type, and molecular subtype.
Figure 4AUC values for prediction models of TP53 and PIK3CA mutation.TP53+/−, TP53 mutation/wild-type; PIK3CA+/−, PIK3CA mutation/wild-type; Asterisk (*) on the segments indicate if a prediction model is significantly higher than another (*p < 0.05; **p < 0.01; ***, p < 0.001; ns, not significant). Note that overall, the combined model for PIK3CA mutation significantly outperforms ultrasound and molecular subtype predictors.
With the cut-off of 0.390 and 0.540 in training and validation sets, the sensitivity and specificity of model 1 were 81%, 43.6%, and 27.1%, 88.1%, respectively. Moreover, result from the Hosmer-Lemeshow test was not significant in the training set (p = 0.992), but it was significant in the validation set (p = 0.006), indicating that the model had a poor fit. Calibration curves of the prediction model of gene mutation revealed good agreement between predictions and observations in the training set (p > 0.05, Fig 5a), while this was poor for the validation set (p < 0.05, Fig 5b).
Figure 5Calibration curves for prediction models targeting TP53 and PIK3CA mutations in the training and validation sets. Calibration curve of the combined model (model 1) for TP53 mutation in the training (a) and validation (b) set. Calibration curve of the combined model (model 2) for PIK3CA mutation in the training (c) and validation (d) set. The solid line represents the ideal reference line that predicted gene mutation corresponds to the actual outcome, the short-dashed line represents the apparent prediction of the combined model, and the long-dashed line represents the ideal estimation. The prediction performance for PIK3CA (model 2) in the training and validation sets show closely observed rates.
In term of PIK3CA mutation, the performances of US (diameter, AUC = 0.610, 95% CI = [0.533-0.681]; type, AUC = 0.553, 95% CI = [0.514-0.591]) and molecular subtype (AUC = 0.642, 95% CI = [0.572-0.712]) were significantly improved by combination of US and molecular subtype (model 2, AUC = 0.741, 95% CI = [0.678-0.804], DeLong test, all p < 0.05, Fig 4a, Table 4) in the training set, suggesting useful discrimination ability [
29
]. Moreover, the performance of US type was improved by model 2 (DeLong test, p < 0.001), but the performance of US diameter and molecular subtype was not significantly increased by model 2 (DeLong test, p = 0.073, p = 0.271, Fig 4b). With the cut-off 0.329 and 0.425 in the training and validation sets, the sensitivity and specificity of model 2 were 80.6%, 58.7%, and 80.6%, 62.1%, respectively. The Hosmer-Lemeshow indicated a good model fit in the training and validation sets (p = 0.631, p = 0.226, respectively). The calibration curve demonstrated good agreement between the predictions and observations in both the training and validation sets (p > 0.05, Fig 5c- d).DISCUSSION
Several studies have reported the association between imaging phenotypes and underlying molecular landscape in tumors (
11
, 12
, 13
, 14
, 15
,- Li H
- Giger ML
- Huynh BQ
- et al.
Deep learning in breast cancer risk assessment: evaluation of convolutional neural networks on a clinical dataset of full-field digital mammograms.
J Med Imaging (Bellingham). 2017; 4041304https://doi.org/10.1117/1.JMI.4.4.041304
20
,21
). Our study showed that calcifications on US was an independent predictor of TP53 mutation, whereas diameter on US and US type were independent predictors of PIK3CA mutation in breast cancer. A combination of US features and molecular subtype exhibited acceptable ability (model 2, AUC = 0.741) to predict PIK3CA mutation compared to US features and molecular subtype alone. Notably, a standardized US imaging protocol was applied to ensure the reliability of the results.Calcification is considered as a malignant mammography feature reflecting pathophysiological microcalcifications that derived from breast tumor cells through epithelial-mesenchymal transition (
30
). Although mammography has shown satisfactory detection of calcifications, however, it is expensive to perform in developing countries and not sensitive to high breast density for Asian women. With the advances in US, most microcalcifications (with a maximum diameter of 1mm) that correlate well with mammography could be detected (31
,32
). We found that the high occurrence of calcifications on US (Fig 6a) was associated with TP53 mutation. This gene controls cell proliferation and is associated with breast cancer aggressiveness (33
), with loss of TP53 leading to epithelial-mesenchymal transition process (34
,35
). These suggested that calcifications on US might associate with cancer progression and metastasis.
Figure 6Illustration of standardized breast US examination based on multiple images. (a) a 65-year-old breast cancer patient with TP53 mutation. Compared with I mage-1, an additional calcification (Arrow) is shown by Image-2 with the multiple-plane US. (b) a 43-year-old patient with PIK3CA mutation. A low echo area looks like geography with indistinct margin (*，Asterisk) is shown in Image-1, which is confirmed by Image-2.
Tumor size reflects tumor proliferation and is an important prognostic indicator of breast cancer (
36
). A rapid growth rate has been characterized as an indicative feature to identify malignant lesions. Our results suggested that diameter on US was an independent predictor of PIK3CA mutation. PIK3CA mutation has direct activation with the PI3K pathway, and activation of this pathway is primarily involved in the development and progression of breast cancer (37
). These might indicate that tumor size on US could imply tumor growth activity. In keeping with our findings, Li et al. (38
) found that larger tumor sizes were more common with PIK3CA mutant carcinomas than wild-type carcinomas, but another study with a larger sample size suggested that small tumor size was associated with PIK3CA mutation (39
). The discrepancy between their findings and those of our study may be attributed to the different sample sizes and statistical manner of reporting tumor size. The tumor size in our study was reported as a numerical variable to improve the precision rather than the categorical variable (40
).Mass and non-mass-like tumor shapes (Fig 6b) are geometric features that describe the roundness of tumor shape. The mass-like tumor was considered to have an infiltrating ability immediately acquired after the tumor formation (
41
). Our study found that mass-like of US type was associated with PIK3CA mutation. This may imply that the mass of US type might be associated with the progression of breast cancer.- Morishima IUE
- Tohno E
- Tsunoda-Shimizu H
- Kujiraoka Y.
- Takasaki M
Ultrasonic diagnosis of non-mass image-forming breast cancer.
in: Ueno E Shiina T Kubota M Sawai K Research and development in breast ultrasound. Springer,
Tokyo2005: 127-134https://doi.org/10.1007/4-431-27008-6_18
The distribution of TP53 and PIK3CA mutations across the molecular subtype were in line with previous studies (
42
,43
), with the majority of TP53 mutation appearing in Luminal B/HER2+, HER2+/ER–, and TN subtypes, and the majority of PIK3CA mutations appearing in Luminal A and Luminal B/HER2- subtypes. Additionally, the molecular subtype was found to be a prime predictor of gene mutation in our model. TP53 mutation in our result was significantly higher than those in The Cancer Genome Atlas (TCGA) database (45.3% and 41.3% in the training and validation set vs. TCGA 37%) (24
). We supposed the disparity of the mutation frequencies of TP53 might result from the lower proportion of patients with ER+/HER- subtype recruited in our cohort compared to the TCGA database. The PIK3CA mutation frequency in our study (38.3% and 33.6% in the training and validation set) was consistent with its high frequency in other publicly available genomic databases (about 36%∼38%) (1
,24
).The model we developed identifying at high risk of PIK3CA mutation may support the clinical decision to benefit from genetic testing. Corresponding to the predictive model 2 with the specificity of 58.7%, the use of the model could have avoided genetic testing in about 60% of the patients with finally confirmed wild-type PIK3CA. Assess of PIK3CA mutations has been included in the National Comprehensive Cancer Network guidelines (Version 2.2022 Invasive Breast Cancer) for HR-positive/HER2-negative breast cancer (
9
). Our results of the high frequency of PIK3CA mutation in Luminal A (33/66, 50.0%) and Luminal B/HER2– (52/103, 50.5%) patients indicated that assessment of genetic testing is necessary for finding already existing PIK3CA mutation in breast cancer. For overall performance, our results indicated that a combination of US and molecular subtypes is a more accurate predictor model for PIK3CA mutation in both sets. The AUC of our work was similar to a previous work in which a gray-level co-occurrence matrix texture feature on breast MRI was used to analyze TP53 and PIK3CA mutant status (16
). However, the molecular subtype was not considered in the study, probably due to the limited sample size. Although the discrimination and application ability of our model needs to be improved, the biological interpretation of US semantical features may give insight to advanced radiomics technology such as machine learning.Despite our efforts to minimize biases, including using a uniform patient population, standardized imaging protocols, and independent training and validation sets, the study still had several limitations. Firstly, the US features were subjective observer-dependent imaging descriptors. However, we applied various efforts, such as using same machine, to achieve concordance, which was confirmed by the results of the ICC test. Secondly, this was a single-center study, multi-institutional datasets are required to validate the applicability of the outcomes herein. Thirdly, the biological context of the informative US features should be explored toward general acceptance of radiomics as a precision diagnostic, predictive, or prognostic method, with validation on the hierarchical diagram of genomics, transcriptomics, proteomics, metabolites, and radiomics.
CONCLUSION
Nevertheless, our findings revealed that US features could reflect the underlying status of gene mutations, thereby laying the groundwork for future establishment of biological meaning to the US radiomics approach. This tool has great potential in both clinical prediction and treatment of diseases.
Acknowledgments
This study was supported by the National Natural Science Foundation of China ( 81830058 , 82071945 ), Joint Fund Project by Shanghai Anticancer Association Fudan University Shanghai Cancer Center ( YJQN202109 ).
Appendix. Supplementary materials
References
- The somatic mutation profiles of 2433 breast cancers refines their genomic and transcriptomic landscapes.Nat Commun. 2016; 7: 11479https://doi.org/10.1038/ncomms11479
- PI3K in cancer: divergent roles of isoforms, modes of activation and therapeutic targeting.Nat Rev Cancer. 2015; 15: 7-24https://doi.org/10.1038/nrc3860
- Recent advances of therapeutic targets based on the molecular signature in breast cancer: genetic mutations and implications for current treatment paradigms.J Hematol Oncol. 2019; 12: 38https://doi.org/10.1186/s13045-019-0725-6
- Mechanistic basis for PI3K inhibitor antitumor activity and adverse reactions in advanced breast cancer.Breast Cancer Res Treat. 2020; 181: 233-248https://doi.org/10.1007/s10549-020-05618-1
- The phosphoinositide 3-kinase pathway.Science. 2002; 296: 1655-1657https://doi.org/10.1126/science.296.5573.1655
- PIK3CA mutation inhibition in hormone receptor-positive breast cancer: time has come.ESMO Open. 2020; 5https://doi.org/10.1136/esmoopen-2020-000890
- 70-gene signature as an aid for treatment decisions in early breast cancer: updated results of the phase 3 randomised MINDACT trial with an exploratory analysis by age.Lancet Oncol. 2021; 22: 476-488https://doi.org/10.1016/S1470-2045(21)00007-3
- Gene expression profiling predicts clinical outcome of breast cancer.Nature. 2002; 415: 530-536https://doi.org/10.1038/415530a
- NCCN Clinical Practice Guidelines in Oncology: Breast Cancer.Version. 2022; 2
- The biological meaning of radiomic features.Radiology. 2021; 299: E256https://doi.org/10.1148/radiol.2021219005
- Somatic mutations drive distinct imaging phenotypes in lung cancer.Cancer Res. 2017; 77: 3922-3930https://doi.org/10.1158/0008-5472.CAN-17-0122
- Prediction of core signaling pathway by using diffusion- and perfusion-based MRI radiomics and next-generation sequencing in isocitrate dehydrogenase wild-type glioblastoma.Radiology. 2020; 294: 388-397https://doi.org/10.1148/radiol.2019190913
- Radiogenomics in head and neck cancer: correlation of radiomic heterogeneity and somatic mutations in TP53, FAT1 and KMT2D.Strahlenther Onkol. 2019; 195: 771-779https://doi.org/10.1007/s00066-019-01478-x
- Deciphering genomic underpinnings of quantitative mri-based radiomic phenotypes of invasive breast carcinoma.Sci Rep. 2015; 5: 17787https://doi.org/10.1038/srep17787
- Deep learning in breast cancer risk assessment: evaluation of convolutional neural networks on a clinical dataset of full-field digital mammograms.J Med Imaging (Bellingham). 2017; 4041304https://doi.org/10.1117/1.JMI.4.4.041304
- Evaluation of TP53/PIK3CA mutations using texture and morphology analysis on breast MR.Magnetic resonance imaging. 2019; 63: 60-69
- Breast ultrasonography: state of the art.Radiology. 2013; 268: 642-659https://doi.org/10.1148/radiol.13121606
- The role of shear wave elastography in differentiating idiopathic granulomatous mastitis from breast cancer.Acad Radiol. 2021; 28: 339-344https://doi.org/10.1016/j.acra.2020.02.008
- Comparison of cancers detected by screening breast ultrasound and digital breast tomosynthesis.Acad Radiol. 2022; 29: 339-347https://doi.org/10.1016/j.acra.2021.01.008
- Radiogenomic analysis of breast cancer by using B-mode and vascular US and RNA sequencing.Radiology. 2020; 295: 24-34https://doi.org/10.1148/radiol.2020191368
- Identifying ultrasound features of positive expression of Ki67 and P53 in breast cancer using radiomics.Asia Pac J Clin Oncol. 2021; 17: e176-e184https://doi.org/10.1111/ajco.13397
- Spectrum of MAP3K1 mutations in breast cancer is luminal subtype-predominant and related to prognosis.Oncol Lett. 2022; 23: 68https://doi.org/10.3892/ol.2022.13187
- GATA3 mutations define a unique subtype of luminal-like breast cancer with improved survival.Cancer. 2014; 120: 1329-1337https://doi.org/10.1002/cncr.28566
- Comprehensive molecular portraits of human breast tumors.Nature. 2012; 490: 61-70https://doi.org/10.1038/nature11412
- Limited family structure and triple-negative breast cancer (TNBC) subtype as predictors of BRCA mutations in a genetic counseling cohort of early-onset sporadic breast cancers.Breast Cancer Res Treat. 2014; 148: 415-421https://doi.org/10.1007/s10549-014-3167-4
- Strategies for subtypes—dealing with the diversity of breast cancer: highlights of the St. Gallen International Expert Consensus on the Primary Therapy of Early Breast Cancer 2011.Ann Oncol. 2011; 22: 1736-1747https://doi.org/10.1093/annonc/mdr304
- Characterization of the genomic landscape and actionable mutations in Chinese breast cancers by clinical sequencing.Nat Commun. 2020; 11: 5679https://doi.org/10.1038/s41467-020-19342-3
- The Danish Medical Statistics and Danish practical research.Ugeskr Laeger. 1990; 152: 828-829
- Measuring the accuracy of diagnostic systems.Science. 1988; 240: 1285-1293https://doi.org/10.1126/science.3287615
- A molecular view of pathological microcalcification in breast cancer.J Mammary Gland Biol Neoplasia. 2016; 21: 25-40https://doi.org/10.1007/s10911-015-9349-9
- Sonographic detection of microcalcifications—potential of new method.Ultraschall Med. 2012; 33: 357-365https://doi.org/10.1055/s-0031-1299128
- Breast cancer detection in a screening population: comparison of digital mammography, computer-aided detection applied to digital mammography and breast ultrasound.J Breast Cancer. 2016; 19: 316-323https://doi.org/10.4048/jbc.2016.19.3.316
- The genomic landscape of metastatic breast cancer highlights changes in mutation and signature frequencies.Nat Genet. 2019; 51: 1450-1458https://doi.org/10.1038/s41588-019-0507-7
- p53 plays a role in mesenchymal differentiation programs, in a cell fate dependent manner.PLOS One. 2008; 3: e3707https://doi.org/10.1371/journal.pone.0003707
- p53 regulates epithelial-mesenchymal transition and stem cell properties through modulating miRNAs.Nat Cell Biol. 2011; 13: 317-323https://doi.org/10.1038/ncb2173
- The 2019 World Health Organization classification of tumors of the breast.Histopathology. 2020; 77: 181-185https://doi.org/10.1111/his.14091
- The PI3K pathway in human disease.Cell. 2017; 170: 605-635https://doi.org/10.1016/j.cell.2017.07.029
- PIK3CA mutations in breast cancer are associated with poor outcome.Breast Cancer Res Treat. 2006; 96: 91-95https://doi.org/10.1007/s10549-005-9048-0
- Tumor PIK3CA genotype and prognosis in early-stage breast cancer: a pooled analysis of individual patient data.J Clin Oncol. 2018; 36: 981-990https://doi.org/10.1200/JCO.2017.74.8301
- Validating a transnational fracture treatment registry using a standardized method.BMC Med Res Methodol. 2019; 19: 241https://doi.org/10.1186/s12874-019-0862-1
- Ultrasonic diagnosis of non-mass image-forming breast cancer.in: Ueno E Shiina T Kubota M Sawai K Research and development in breast ultrasound. Springer, Tokyo2005: 127-134https://doi.org/10.1007/4-431-27008-6_18
- The sgenomic landscape of breast cancer as a therapeutic roadmap.Cancer Discov. 2013; 3: 27-34https://doi.org/10.1158/2159-8290.CD-12-0462
- TP53 mutation spectrum in breast cancer is subtype specific and has distinct prognostic relevance.Clin Cancer Res. 2014; 20: 3569-3580https://doi.org/10.1158/1078-0432.CCR-13-2943
Article Info
Publication History
Published online: April 18, 2022
Accepted:
February 23,
2022
Received in revised form:
February 23,
2022
Received:
February 5,
2022
Identification
Copyright
© 2022 The Association of University Radiologists. Published by Elsevier Inc.
User License
Creative Commons Attribution – NonCommercial – NoDerivs (CC BY-NC-ND 4.0) | How you can reuse 
Elsevier's open access license policy
Creative Commons Attribution – NonCommercial – NoDerivs (CC BY-NC-ND 4.0)
Permitted
For non-commercial purposes:
- Read, print & download
- Redistribute or republish the final article
- Text & data mine
- Translate the article (private use only, not for distribution)
- Reuse portions or extracts from the article in other works
Not Permitted
- Sell or re-use for commercial purposes
- Distribute translations or adaptations of the article
Elsevier's open access license policy
00131-3&id=gr1.jpg){kind=link}
00131-3&id=gr2.jpg){kind=link}
00131-3&id=gr3.jpg){kind=link}
00131-3&id=gr4.jpg){kind=link}
00131-3&id=gr5.jpg){kind=link}
00131-3&id=gr6.jpg){kind=link}