AI-enhanced simultaneous multiparametric ¹⁸F-FDG PET/MRI for accurate breast cancer diagnosis

This prospective single-institution study was approved by the institutional review board, and written informed consent was obtained from all participants. From June 2016 to July 2020, 154 patients were included in the study and underwent simultaneous multiparametric ¹⁸F-FDG PET/MRI of the breast for diagnostic purposes. Patients were included according to the following inclusion criteria: > 18 years of age; not pregnant nor breastfeeding; and imaging abnormality (i.e., Breast Imaging-Reporting and Data System (BI-RADS) 0, 4/5) on ultrasound and/or mammography (i.e., asymmetries, microcalcifications, architectural distortion, breast mass). Exclusion criteria were no histopathology or follow-up available; incomplete ¹⁸F-FDG PET/MRI examinations; ¹⁸F-FDG PET/MRI images not suitable for subsequent quantitative and radiomics analysis (e.g., image artifacts, incomplete dynamic scans); previous treatments; and contraindications for MRI examination. Thus, 102 patients (mean age 50 years, age range 23–82 years) with 120 breast lesions (101 malignant and 19 benign) were finally included in this study. The BI-RADS category distribution of included lesions was BI-RADS 0 (n = 8), BI-RADS 4 (n = 16), and BI-RADS 5 (n = 96). The flowchart of the patient selection process is given in Fig. 1.

Histology was used as the reference standard for lesions classified as BI-RADS 4 (n = 22) or 5 (n = 95) at ¹⁸F-FDG PET/MRI. In patients with malignant lesions, the reference standard was histological analysis of the surgical specimen; in patients who received neoadjuvant treatment, the biopsy results were considered the reference standard. In three lesions classified as BI-RADS 2 (n = 2) or BI-RADS 3 (n = 1) at ¹⁸F-FDG PET/MRI, stable imaging follow-up was available for at least 2 years.

All patients underwent simultaneous multiparametric ¹⁸F-FDG PET/MRI performed using a Biograph mMR system (Siemens, Germany), which is an MRI-compatible PET detector integrated with a 3.0 MRI scanner [16].

Patients fasted at least 5 h before receiving an intravenous application of ¹⁸F-FDG at a dose of 2.5–3.5 MBq/kg body weight. All measured blood glucose levels were less than 150 mg/dL (8.3 mmol/L) prior to tracer injection. The PET/MRI acquisition started after an uptake time of 60 min. MRI-based attenuation correction was performed using the standard Dixon-based attenuation correction method [17, 18]. A three-dimensional (3D) acquisition technique was used that offered an axial field of view (FOV) of approximately 26 cm and a transverse FOV of 59 cm with a sensitivity of 13.2 cps/kBq. Data acquisition was done for 30 min. Static PET images were reconstructed using ordinary Poisson 3D ordered subset expectation maximization (OP-OSEM) (with Gaussian scatter correction) with 3 iterations and 21 subsets into a 172-zoom 1.0 image matrix including all standard corrections (normalization, scatter, random coincidences, and decay).

Multiparametric MRI was performed using a dedicated 16-channel breast coil (Rapid Biomedical, Germany), and the imaging protocol consisted of the following sequences:

Two board-certified radiologists with 10 and 6 years of experience in breast imaging independently evaluated MRI data. A nuclear medicine physician with 10 years of experience and a radiologist with 6 years of experience who was trained in hybrid imaging under the supervision of a nuclear medicine physician independently evaluated PET images. Readers were blinded from final histopathological results and previous examinations. To assess the intraobserver reproducibility of PET/MRI quantitative parameter measurements, all lesions were reassessed by the same readers after a washout period of 4 weeks.

PET/MRI images were imported to dedicated software (ITK-SNAP v. 3.6.0) [22] for lesion segmentation. A radiologist with 6 years of experience in breast imaging annotated each lesion on DCE, DWI, PET, and T2-weighted images. First, whole breast lesions were segmented on DCE-MR images using a semi-automated method. The second post-contrast time point was chosen for lesion segmentation, in order to better depict tumor enhancement compared to the surrounding breast parenchyma. The same approach was applied to DWI and PET images. Finally, manual segmentation was performed to annotate breast lesions on T2-weighted images slice by slice. In all steps, care was taken to avoid the inclusion of cystic/necrotic areas. When a biopsy marker was present, a distance of at least 2 mm was kept. Examples of tumor segmentation are given in Fig. 3.

Considering the unbalanced distribution of benign and malignant breast lesions, adaptive synthetic sampling was employed to equalize class sizes [23]. Data for all four image types was initially reduced to 16 grey levels. Radiomics features were calculated using the Computational Environment for Radiological Research (CERR) [24]. DCE, T2-weighted, ADC, and PET images were used for radiomics feature extraction. Segmentations performed on DWI images were used for the extraction of radiomics features from ADC images. Considering that T2-weighted and ADC images were not isotropic, feature extraction was performed in a 2D fashion for each slice and then aggregated over the whole lesion (BTW3 as defined by the Image Biomarker Standardisation Initiative) [25]. Least Absolute Shrinkage and Selection Operator (LASSO) regression was then utilized to determine which radiomics features were of most importance. LASSO forces the sum of the regression coefficients to be less than a fixed value, which in turn forces certain coefficients to be zero, thus excluding them from affecting prediction. For this work, a maximum of the top 5 most important features were selected, to avoid overfitting the limited datasets available. LASSO was employed due to its fast nature, its ability to avoid overfitting, and the fact that it can be applied even when the number of features is greater than the number of cases/samples [26, 27]. Diagnostic models were then developed in MATLAB using a fine Gaussian support vector machine (SVM), one of the most employed machine learning (ML) classifiers in medical imaging [28]. Since there is a short supply of data to enable a split into training, validation, and test sets, the selection of ML methodology becomes especially important. An SVM was utilized since it is known to work well for small datasets, the resulting models are memory efficient (since only the coefficients corresponding to the support vectors are non-zero), they can solve both linear and non-linear problems, and they usually provide good performance [29, 30]. An SVM algorithm works by creating a hyperplane which separates the data into the desired classes. Again, since there is insufficient data to split into traditional training and test sets, fivefold cross-validation was employed since it gives the model the opportunity to train on multiple train-test splits. This results in a better indication of how well the model will perform on unseen data. Data were initially standardized (z-score calculation with mean 0 and standard deviation 1) to prevent dependence on any individual parameter, especially those which contain high values. This process was then repeated 1000 times to provide final diagnostic metrics. Analysis was performed for each of the four image types independently and then in various combinations to assess potential improvements in diagnostic accuracy for the discrimination of benign and malignant breast lesion.

DCE-MRI was assessed according to the fifth edition of the BI-RADS lexicon [31]. A BI-RADS category from 2 to 5 was assigned to each lesion. BI-RADS scores were then dichotomized as follows: 2–3 = benign and 4–5 = malignant. Subsequently, ADC values were calculated for each lesion, as described above. An ADC value of 1.3 × 10⁻³ mm²/s was used as the diagnostic threshold for defining benignity and malignancy, as suggested by the European Society of Breast Imaging (EUSOBI) consensus statement on DW imaging [21]. Lesions showing ADC values equal to or greater than 1.3 × 10⁻³ mm²/s were classified as benign, while lesions with ADC values lower than 1.3 × 10⁻³ mm²/s were classified as malignant. On PET, a lesion was classified as benign if it did not show ¹⁸F-FDG uptake higher than the above background activity; conversely, a lesion showing ¹⁸F-FDG uptake greater than the surrounding parenchyma was classified as malignant [32]. To achieve a final diagnosis, the following criteria were applied for the combined DCE-MRI, DWI, and PET evaluation:

Intra- and interobserver reproducibility of quantitative parameter measurements was assessed using intraclass correlation coefficient (ICC) analysis. Agreement was rated as follows: poor when ICC is less than 0.5, moderate when ranging from 0.5 to 0.75, good when ranging from 0.75 to 0.90, and excellent when greater than 0.90 [33]. The Kolmogorov–Smirnov test was performed to assess whether quantitative parameters were distributed normally. The independent t-test or Mann–Whitney U test was used to compare quantitative parameters between benign and malignant breast lesions. Diagnostic accuracy, sensitivity, specificity, and positive and negative likelihood ratio of the radiologists and nuclear medicine physician’s performance in classifying breast lesions were also calculated. Receiver operating characteristic (ROC) curves of the BI-RADS score as well as significant quantitative DWI, perfusion, and PET parameters for breast cancer diagnosis were also calculated. Differences in terms of performance between the different radiomics models as well as between the best performing radiomics model and clinical interpretation were assessed using McNemar’s test. A p value ≤ 0.05 was considered statistically significant. Statistical analysis was performed using SPSS, version 25.0, which was released in 2017 (Armonk, NY: IBM Corp) and MedCalc 18 (MedCalc software bvba).

Of the 120 included breast lesions, 101 (84%) were malignant and 19 (16%) were benign. Histological features of included breast lesions are reported in Tables 1 and 2.

Breast carcinomas showed significantly higher maximum diameter (28.5 vs 10.68 mm, p = 0.035) and SUVmax (6.22 vs 3.09, p = 0.003) as well as lower ADCmean (1.42 vs 9.23 × 10⁻³ mm²/s, p < 0.001) and MTT (77.17 vs 118.90 s, p < 0.001) than benign breast lesions. Mean values of quantitative parameters in both benign and malignant breast lesions with corresponding significance levels are reported in Supplementary Material Table S1. ICC values of all quantitative parameters including intra- and interobserver reproducibility are shown in Supplementary Material Table S2. Intra- and interobserver reproducibility of quantitative parameter measurement ranged from moderate to excellent.

A total of 101radiomic features were extracted in six classes (22 first order, 26 based on grey-level co-occurrence matrices, 16 based on run length matrices, 16 based on size zone matrices, 16 based on neighborhood grey-level dependence matrices, and five based on neighborhood grey tone difference matrices) from DCE, ADC, T2-weighted, and PET images, respectively. Eight radiomics models were developed to predict breast cancer diagnosis, based on different combinations of multiparametric ¹⁸F-FDG PET/MRI images. Radiomics models with corresponding selected radiomics features are reported in Table 3. Firstly, a radiomics model based on quantitative parameters alone was built. ADCmean of breast lesions, MTT, and SUVmax were selected by the LASSO regression and used by the SVM classifier, obtaining an area under the curve (AUC) of 0.981 for correctly classifying breast lesions.

Thereafter, the accuracy of diagnostic models based on radiomics features extracted from individual DCE, T2-weighted, ADC, and PET images was explored. Among these models, the best performance in discriminating between benign and malignant breast lesions was obtained by an SVM classifier using features extracted from ADC images (AUC 0.937, 95% confidence interval (CI): 0.901–0.973). The model based on T2-weighted features performed worse, with an AUC of 0.793 (95% CI: 0.732–0.855). Based on these findings, two radiomics models were built combining (1) radiomics features extracted from ADC maps and DCE images, and (2) radiomics features extracted from ADC, DCE, and PET images. Of these, the latter showed the best performance for breast cancer diagnosis (AUC 0.969, 95% CI: 0.947–0.990). Finally, an integrated model combining quantitative parameters and radiomics features extracted from DCE, PET images, and ADC maps was built. MTT and ADCmean of breast lesions and radiomics features extracted from ADC maps and PET images were selected by LASSO regression. This model obtained the highest accuracy in discriminating between benign and malignant breast lesions, with an AUC of 0.983 (95% CI: 0.962–1.000). A summary of all radiomics models with corresponding accuracy metrics, including area under the receiving operating characteristic curve (AUROC), diagnostic accuracy, sensitivity, specificity, and positive and negative likelihood ratio is reported in Table 4.

Using McNemar’s test (Table 5), the performance of the integrated model combining quantitative parameters and radiomics features was higher but not significantly different from that of the other radiomics models (p > 0.069).

The distribution of BI-RADS descriptors, ADC values, and PET findings in the study population is reported in Supplementary Material Table S3. Clinical interpretation achieved a diagnostic accuracy of 0.958 (95% CI: 0.905–0.986), sensitivity and specificity of 100% (95% CI: 96–100%) and 73.7% (95% CI: 49–91%), positive and negative likelihood ratio of 3.80 (95% CI: 1.79–8.06), and 0 in discriminating between benign and malignant breast lesions.

All radiomics models but the one based on T2-weighted radiomics features achieved greater AUCs than clinical interpretation of ¹⁸F-FDG PET/MRI. However, this difference was not significant (p = 0.508). ROC curves of BI-RADS scores along with significant quantitative DWI (ADCmean), perfusion (MTT) and PET (SUVmax) parameters for discriminating between benign and malignant breast lesions are illustrated in Supplementary Material Figure S1. Accuracy metrics of clinical interpretation, including all imaging modalities and the combined evaluation, as well as AI assessment, are summarized in Supplementary Material Table S4.