inating different types of breast cancer167. Our hypothesis is that each confounding class
requires a distinct set of features to be able to discriminate it from prostate cancer successfully and that this pair-wise classification approach will yield improved discriminability compared to a monolithic classifier aempting to distinguish cancer from all benign classes simultane- ously.
4.2 Materials and methods
4.2.1 Patient data
Pre-operative multi-parametric MRIs and radical prostatectomy specimens were included ret- rospectively for 70 patients at the Radboud University Nijmegen Medical Centre. MRIs were acquired between January 1st 2009 and June 1st 2013. e Institutional Review Board waived the need for informed consent.
4.2.2 MRI acquisition
MRI acquisition was performed using a 3 Tesla MRI scanner (either a TrioTim or a Skyra; Siemens, Erlangen, Germany). Cases were acquired both with and without an endorectal coil. A pelvic phased array coil was always used. e multi-parametric protocol consisted of three T2-weighted images in orthogonal directions, diffusion-weighted imaging (three b-values av- eraged over three orthogonal directions, 50, 400-500 and 800) and dynamic contrast-enhanced imaging (15 mL of Dotarem; Guerbet, France). e transversal T2-weighted images were acquired perpendicular to the rectal wall, the diffusion-weighted imaging and the dynamic contrast-enhanced imaging were acquired in the same orientation. Further acquisition details can be found in Table 4.1.
4.2.3 Prostatectomy slide selection and annotation
staining the specimens were evaluated by one expert urological pathologist (C.H.v.d.K, with 20 years of experience). e pathology slides were cut in the same orientation as the acquisition of the transversal MRI to remove angulation errors in subsequent registration steps. Tumors were outlined on the microscopic slides and subsequently transferred to the macroscopic pho- tographs of the specimens.
e H&E stained slide containing the tumor with the highest Gleason score was selected to be digitized using a digital slide scanner (VS120-S5, Olympus, Japan) at 10x or 20x, cor- responding to a resolution of 0.6 um and 0.3 um respectively. If multiple slices contained a tumor with the same Gleason grades, the slice with the largest tumor volume was digitized. Approximately half of the specimens were whole-mount slides, the other half consisted of parts (usually two or four). In case the specimen consisted of parts, all parts belonging to
PS SR ST NS ET RT FA Other T2W Turbo spin- echo 0.28 – 0.6 mm 3.0 – 4.0 mm 13 - 19 101 – 104 ms 3540 – 6840 ms
120 - 160 Acquired in three orthogo- nal directions: transversal, sagial and coronal. DWI Echo pla-
nar 1.6 - 2 mm 3 mm - 4 mm 15 - 20 61 – 81 ms 2300 – 3600 ms 90 3 b-values: 50, 400 – 500, 800 averaged over 3 directions. Apparent diffusion coeffi- cient map calculated by the scanner soware. Some scans also include b-value 0. DCE Turbo Flash 1.5 – 1.8 mm 3.2 – 5 mm 12 - 15 1.41 - 1.47 ms 36 ms 10 - 14 Temporal resolution of 3.38 – 4.65 seconds, 36 – 50 timepoints. 15 mL contrast agent used (Dotarem, Guer- bet, France)
Table 4.1: MRI sequence details for the different types of acquisitions. PS = pulse sequence, SR = spatial resolution, ST = slice thickness, NS = number of slices, ET = echo time, RT = repetition time, FA = flip angle.
one slide were digitized. Aer digitization the digital slides were annotated using the Aperio ImageScope soware (Aperio, USA) for the presence of cancer, BPH, PIN, atrophy or inflam- mation by one of two urological pathologists (N.S. with 8 years of experience or R.E. with 7 years of experience).
4.2.4 Co-registration of prostatectomy specimens and MRI
To map the annotations on the histopathology sections to the corresponding MRI sections, the MRI and the pathology slide have to be registered. First, the slice in the MRI corresponding to the prostatectomy slide has to be established168. e number of slices in the MRI the prostate
was visible on were counted. Subsequently, the number of slides in the prostatectomy was counted. Using the number of the prostatectomy slide the most likely corresponding MRI slice is then given as:
SMR= TMR TP
SP (4.1)
whereSMR is the slice number in the MRI,TMR the total number of prostate slices in the MRI,
TP the total number of prostate slices in pathology and SP the slice number of the selected
pathology slice. is is similar to the approach presented by Hambrock et al.51. e selected
MR and pathology slices where subsequently visually assessed for correspondence by a med- ical imaging researcher (G.L., four years of experience with prostate MRI) and corrected if deemed necessary. Aer establishing the corresponding slice it was registered to the MRI us- ing an interactive b-spline elastic registration method, which has successfully been applied in
4.2 Materials and methods 73 a number of studies94,95. To drive the registration corresponding points on the boundary of the
prostate on the MRI and the pathology were selected by a medical imaging researcher (G.L., four years of experience with prostate MRI). Aer the corresponding points were established, the registration algorithm mapped the prostatectomy slide and the annotations to the corre- sponding MRI section. During selection of the boundary points, the researcher was blinded to the pathology annotations. An example of the process is illustrated in Figure 4.1.
4.2.5 Computer-extracted features
Following co-registration, a number of MRI and computer-extracted features were obtained from within the regions corresponding to the cancer, BPH, PIN, atrophy and inflammation. To obtain a single feature vector per region of interest (ROI) mapped onto the MRI, the median value of each feature across the voxels within the ROI is calculated. All features are calculated in 2D, as we register a single prostatectomy slide to the MRI, resulting in 2D annotations. A listing of these features and their associated descriptions can be found in Table 4.2.
Category Feature name Calculated on Parameters
Intensity T2Ws169 T2W (Transversal) ADC DWI b800 DWI Texture 2D multi-scale Gaussian derivatives T2Ws Up to 2nd order,σ=2.0, 2.7, 4.1 and 6.0mm
2D multi-angle Gabor T2Ws θ=0, π4, π2, 34π. λ=2, 3 and 4
mm 2D Li multi-scale blob-
ness170
T2Ws, ADC, b800, Ktrans, kep,ve, time-to-peak, max-
imum enhancement, wash- out rate
σ=2.0, 2.7, 4.1 and 6mm
Pharmacokinetic
Time-to-peak55 DCE Maximum enhancement55 DCE Wash-out rate55 DCE
Ktrans 57 DCE
ve57 DCE
kep57 DCE
MR intensity features
MR intensity features are extracted from the transversal T2-weighted image volume and the diffusion-weighted imaging. In T2-weighted imaging the non-standardness of the MR intensi- ties, especially between endorectal coil and non-endorectal coil cases, can cause problems for quantitative computerized analysis. As such we developed a method which uses MR pulse se- quence equations, a proton-density-weighted image and an automatically segmented muscle
ROI to remove most of the T2-weighted intensity non-standardness169.
In addition to the standardized T2-intensity we included the apparent diffusion coefficient (ADC) as a feature in combination with the image intensity of the b800 image.
Texture features
We calculated several popular texture filters, namely Gaussian derivatives and Gabor features, which have shown to be successful in discriminating prostate cancer from other tissue in previ- ous studies94,95. To make sure these feature do not also suffer from intensity non-standardness
we calculated them on the standardized T2W image.
Furthermore, to assess the focality of lesion appearance on the different MRI parameters,
several blobness features were calculated using the techniques presented by Li et al170. Pa-
rameter seings for these features are listed in Table 4.2.
Pharmacokinetic features
DCE MRI has been shown to differentiate between inflammation and prostate cancer relative to normal tissue162. In clinical diagnosis oen the shape of the enhancement curve is used to
assess lesion malignancy. However, several groups have developed methods to more quanti- tatively evaluate the tissue curves, including pharmacokinetic modelling55,162,171. We use the
methods presented by Huisman et al.55and Vos et al.57to calculate pharmacokinetic features.
4.2.6 Feature selection and classification
We used sequential forward floating feature selection (SFFS,172) in combination with a linear
discriminant classifier to assess the most discriminative features. SFFS is a feature selection technique in which at each step one feature is added or removed based on a performance met- ric; we used the area under the receiver-operating characteristic curve (AUC). In our setup we force the feature selection to find the 5 most relevant features for each pair-wise classification task (cancer vs. BPH, atrophy, inflammation and PIN, respectively).
We repeated the SFFS procedure to investigate whether the selected features are influenced by cancer grade. We specifically looked at intermediate- and high-grade cancer. Intermediate- grade cancer was defined as cancer with a Gleason grade 3+4 and high-grade cancer was defined as any cancer with a major 4 or any 5 component.
4.3 Results 75