Computational Detection of Extraprostatic Extension of Prostate Cancer on Multiparametric MRI Using Deep Learning

Abstract

The localization of extraprostatic extension (EPE), i.e., local spread of prostate cancer beyond the prostate capsular boundary, is important for risk stratification and surgical planning. However, the sensitivity of EPE detection by radiologists on MRI is low (57% on average). In this paper, we propose a method for computational detection of EPE on multiparametric MRI using deep learning. Ground truth labels of cancers and EPE were obtained in 123 patients (38 with EPE) by registering pre-surgical MRI with whole-mount digital histopathology images from radical prostatectomy. Our approach has two stages. First, we trained deep learning models using the MRI as input to generate cancer probability maps both inside and outside the prostate. Second, we built an image post-processing pipeline that generates predictions for EPE location based on the cancer probability maps and clinical knowledge. We used five-fold cross-validation to train our approach using data from 74 patients and tested it using data from an independent set of 49 patients. We compared two deep learning models for cancer detection: (i) UNet and (ii) the Correlated Signature Network for Indolent and Aggressive prostate cancer detection (CorrSigNIA). The best end-to-end model for EPE detection, which we call EPENet, was based on the CorrSigNIA cancer detection model. EPENet was successful at detecting cancers with extraprostatic extension, achieving a mean area under the receiver operator characteristic curve of 0.72 at the patient-level. On the test set, EPENet had 80.0% sensitivity and 28.2% specificity at the patient-level compared to 50.0% sensitivity and 76.9% specificity for the radiologists. To account for spatial location of predictions during evaluation, we also computed results at the sextant-level, where the prostate was divided into sextants according to standard systematic 12-core biopsy procedure. At the sextant-level, EPENet achieved mean sensitivity 61.1% and mean specificity 58.3%. Our approach has the potential to provide the location of extraprostatic extension using MRI alone, thus serving as an independent diagnostic aid to radiologists and facilitating treatment planning.

Keywords: computer-aided diagnosis; deep learning; extraprostatic extension.

Figure A1

Visualization of predictions for extraprostatic extension in four example cases from the test set; each row shows the leading slice from a different case. Column (1) input T2w images; Column (2) input ADC images; yellow labels outline cancer within the prostate, white labels outline extraprostatic extension (EPE); white arrows point to the small EPE regions; the panels in the corners show zoomed-in 5 mm× 5 mm areas containing EPE. Column (3) EPENet predictions. Predictions are probability maps displayed in a cold to hot color scheme (dark blue–0, dark red–1). Note that EPENet predictions highlight lesions with extraprostatic extension. EPENet does not show any predictions for cancer lesions fully contained within the prostate capsule.

Figure A2

True and false predictions, with ground truth information about cancer and EPE lesion volumes. The x-axis shows the volume of cancer inside the prostate as a percentage of the total prostate gland volume. The y-axis shows the volume of extraprostatic extension as a percentage of the total prostate gland volume. Cases which are ground truth negative for EPE therefore have coordinate $y=0$. (a) All training set cases. (b) Training set, zoomed-in view on the cases with small cancer and EPE volumes. (c) All test set cases. (d) Test set cases, zoomed-in view on the cases with small cancer and EPE volumes.

Figure A2

True and false predictions, with ground truth information about cancer and EPE lesion volumes. The x-axis shows the volume of cancer inside the prostate as a percentage of the total prostate gland volume. The y-axis shows the volume of extraprostatic extension as a percentage of the total prostate gland volume. Cases which are ground truth negative for EPE therefore have coordinate $y=0$. (a) All training set cases. (b) Training set, zoomed-in view on the cases with small cancer and EPE volumes. (c) All test set cases. (d) Test set cases, zoomed-in view on the cases with small cancer and EPE volumes.

Figure A3

True positive (TP) and false positive (FP) predictions, with information about the EPENet prediction volumes. The x-axis shows the total volume of EPENet prediction as a percentage of the prostate gland volume. The y-axis shows the percentage of the prediction volume that is outside the prostate. The TP and FP classes are not linearly separable based on size and position of prediction alone. (a) An example of cut-offs we might select in the training set. (b) Effect of applying training set cut-offs on the test set. (a) Training set cases. The dotted lines show an example of cut-off values we could apply to eliminate some false positives (i.e., declare predictions that have either $x>60\%$ or $y>60\%$ negative for EPE). This would eliminate 14 false positives and create 2 false negatives (15% reduction in FP and 7% reduction in TP). (b) Test set cases, with the example cut-off values from the training set. We observe that applying the cut-offs would eliminate five false positives, but would also lead to three false negatives (14% reduction in FP and 25% reduction in TP).

Figure 1

Flowchart describing our method for generating spatial predictions of cancer lesions with extraprostatic extension. A convolutional neural network takes in T2w and ADC images and outputs a probability map for cancer. We apply post-processing steps of masking, thresholding and connected components to the probability map to obtain a set of lesion candidates. We check each lesion candidate against a set of heuristic rules and determine which are the lesions with suspected extraprostatic extension.

Figure 2

Row 1: Registered MRI and histopathology slices, along with output from cancer detection deep learning model in a patient with bulky extraprostatic extension. (a) T2w image, (b) ADC image, (c) Histopathology image overlaid with ground truth cancer labels and prostate segmentation mask. Purple contour shows prostate gland segmentation, orange contour is cancer inside the prostate, white contour is extraprostatic extension (EPE). (d) Cancer probability map output by the pre-trained deep learning cancer detection model. Row 2: Post-processing steps to generate final predictions for cancer with extraprostatic extension. (e) Cancer probability map after applying dilated prostate mask. (f) Candidate lesions (shown in yellow) were obtained by applying binary threshold and detecting connected components. (g) Candidate lesions were pruned based on their location relative to the prostate; components fully inside or fully outside the prostate were rejected (brown), those crossing the border were accepted (green). For the green lesion candidate, the tumor-capsule contact line (TCL) is displayed in pink. (h) Final prediction map for lesion with EPE.

Figure 3

Grid search over the tumor-capsule contact line length threshold parameter ($TC{L}_{th}$) for EPENet and UNet_EPE. The two panels correspond to the evaluation metric used (displayed on the y axis): sextant-level AUC in panel (a) and patient-level AUC in panel (b). The x axis shows the values of $TC{L}_{th}$ in millimeters. Results are averages computed over the validation sets of the five cross-validation data folds.

Figure 4

EPENet predictions for cancers with extraprostatic extension in one example case from the test set, from apex to base. EPENet does not show any predictions for cancer lesions fully contained within the prostate capsule. Column (1) Schematic illustration of the prostate division into sextant regions. Column (2) Input T2w images. Column (3) Input ADC images; yellow labels outline cancer within the prostate, white labels outline extraprostatic extension (EPE); white arrows point to the small EPE regions; the panels in the corners show zoomed-in 5 mm × 5mm areas containing EPE. This patient shows cancer on all slices, but EPE is present only in rows 2 and 3. Column (4) and (5) show EPENet predictions for different values of the $\alpha$ parameter. Predictions are probability maps displayed in a cold to hot color scheme (dark blue–0, dark red–1).

Figure 5

Receiver operating characteristic curves based on performance metrics at the sextant-level (a) and patient-level (b). Dashed lines are the mean performance over the five cross-validation folds and the shaded regions represent one standard deviation around the average. Solid lines are the ROC results in the test set. EPENet model is shown in blue, UNet_EPE model is green. In panel (b), the star marker represents radiologists performance on the test set.