Dataset characteristics
Two public dataset were used: ‘QIN Breast DCE-MRI’ and ‘QIN-Breast’.
The public dataset ‘QIN Breast DCE-MRI’ from The Cancer Imaging Archive (TCIA) collection [20, 21] is composed of ten patients subjected to DCE-MRI using a Siemens 3-T TIM Trio system with the body coil and a four-channel bilateral phased-array breast coil. Axial bilateral DCE-MRI images with fat saturation and full breast coverage were acquired with a three-dimensional gradient echo-based time-resolved angiography with stochastic trajectories sequence. DCE-MRI acquisition parameters included echo time 2.9 ms and repetition time 6.2 ms; field of view 30–34 cm, in-plane matrix size 320 × 320; and slice thickness 1.4 mm. The total acquisition time was about 10 min for 32–34 image volume sets of 112–120 slices each, with a temporal resolution of 18–20 s. The contrast agent was Gd-HP-DO3A, gadoteridol (Bracco Imaging, Milan, Italy), intravenously injected (0.1 mmol/kg at 2 mL/s) by a programmable power injector timed to commence after acquisition of two baseline image volumes, followed by a 20-mL saline flush. The public data set can be downloaded at https://wiki.cancerimagingarchive.net/display/Public/QIN+Breast+DCE-MRI.
The public dataset ‘QIN-Breast’ from The Cancer Imaging Archive (TCIA) collection [21, 22] is composed of 35 patients subjected to DCE-MRI using a 3-T Philips Achieva system using a dedicated 16-channel bilateral breast coil. Axial bilateral DCE-MRI images with fat saturation and full breast coverage were acquired with a radiofrequency spoiled three-dimensional gradient echo sequence. Acquisition parameters included echo time 7.9 and repetition time 4.6 ms, field of view 22 cm2, in-plane matrix size 192 × 192, and slice thickness 5 mm. For the DCE study, each 20-slice set was collected in 16 s at 25 time points for just under 7 min of dynamic scanning. The contrast agent was Gd-DTPA and gadopentetate dimeglumine (Bayer Health Care Pharmaceuticals, Wayne, NJ, USA) was intravenously injected (0.1 mmol/kg at 2 mL/s) by a programmable power injector timed to commence after acquisition of two baseline image volumes, followed by a 20-mL saline flush. The public data set can be downloaded at https://wiki.cancerimagingarchive.net/display/Public/QIN-Breast.
The NAT protocol administered to these patients was left to the discretion of the treating oncologist based on patient factors such as menopausal status and age as well as tumour characteristics, including size, grade, nodal status and receptor status and was reported by Li et al. in [22]. Both these collections of breast DCE-MRI data contain images from two studies to assess NAT response. Images were acquired at two time points: before and after the first cycle of treatment.
Data analysis
Manual segmentation was performed by an expert breast radiologist (with a 25-year experience) on the post-contrast arterial phase images, drawing manually each slice to obtain the delineating of the whole tumour contours (volume of interest).
Textural features
We considered 50 textural features, including both first-order features (mean, mode, median, standard deviation [SD], median absolute deviation, range (absolute difference between maximum and minimum values), kurtosis, skewness, and interquartile range) and second-order features. Calculations were performed using the ‘TextureToolbox’ of MATLAB R2007a (MathWorks, Natick, MA, USA) that performs texture analysis from an input by region or volume of interest. In particular, this texture analysis package allows for wavelet band-pass filtering, isotropic resampling, discretisation length corrections and different quantitation tools. A detailed description has been provided by Vallières et al. [23]. The toolbox can be downloaded at https://it.mathworks.com/matlabcentral/fileexchange/51948-radiomics. The definition of significant textural features reported in the “Results” section is provided in Additional file 1.
Semiquantitative dynamic parameters
A time-intensity curve can be subdivided into three regions. The first one represents the contrast medium time needed to reach the lesion, and the signal intensity is equal to the basal level before contrast agent injection; the second one shows the increase in signal intensity because of contrast medium absorption (washin) according to the tumour biology; the third one mainly represents the backflow of the contrast medium into the plasma (washout). To estimate shape descriptors, a piecewise linear fitting was made and ten semiquantitative dynamic features described in the literature [24,25,26] were extracted using the approach reported in a previous publication from our group [25], maximum signal difference (MSD), time to peak between washin and washout segments, washin slope (WIS), washout slope (WOS), washin intercept, washout intercept, area under the curve of washin, area under the curve of washout, and area under the curve of washin and washout. The last semiquantitative dynamic feature was the SIS obtained combining linearly the percentage change of MSD and WOS. Therefore, for SIS calculation, the percentage change of MSD [ΔMSD = (MSD1 - MSD2)/MSD1 × 100], and of WOS [ΔWOS = (WOS1 - WOS2)/WOS1 × 100] and their combination as previously described [11] was evaluated. Standardised SIS was given by the following linear combination: 0.7780*ΔMSD + 0.6157*ΔWOS. In order to evaluate the SIS, an OsiriX (Pixmeo SARL, Geneva, Switzerland) plugin has been developed by the authors.
Reference standard and pathological methods
The reference standard was the pathology from surgical specimen. Fifteen pCR patients and 30 non-pCR patients were included in this retrospective study. The pCR was classified according to Miller-Payne grade: grade 1 for no reduction, grade 2 for minor loss (≤ 30%), grade 3 for loss from 30 to 90%, grade 4 for marked loss (> 90%), and grade 5 for no residual invasive cancer. Patients with grades 1, 2, 3, or 4 were scored as non-pCR.
Statistical analysis
Median, SD, and range were calculated as representative values of segmented volumes of interest. Percentage change of median values of parameters obtained before and after the first cycle of treatment was calculated. Receiver operating characteristic analysis was used for obtaining the area under the curve (ROC-AUC). Sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and accuracy were obtained considering the optimal cutoff values identified maximising the Youden index.
For two-group comparisons, we used the non-parametric Kruskal-Wallis test for continuous variables. A p value < 0.05 was considered as significant for univariate analysis. Bonferroni correction was applied for multiple comparisons.
Calculations were performed using the Statistics and Machine Learning Toolbox of MATLAB R2007a (MathWorks, Natick, USA).