1846

mpMRI-based Tumor Probability Maps for Guidance of Targeted Prostate Biopsies
Gabriel Addio Nketiah1, Nienke Bakx1,2, Kirsten Margrete Selnæs 1,3, Adrian Lazaro Breto4, Radka Stoyanova 4,5, Mattijs Elschot 1, and Tone Frost Bathen1

1Department of Circulation and Medical Imaging, Norwegian University of Science and Technology, Trondheim, Norway, 2Department of Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands, 3Clinic of Radiology and Nuclear Medicine, St. Olavs University Hospital, Trondheim, Norway, 4Sylvester Comprehensive Cancer Center, University of Miami, Miami, FL, United States, 5Department of Radiation Oncology, University of Miami, Miami, FL, United States

Synopsis

Despite the improvement offered by the integration of multiparametric magnetic resonance imaging (mpMRI) in biopsy acquisition for prostate cancer diagnosis, the number of negative biopsies remains high with increasing risk of post-biopsy infection and complications. We evaluated the utility of machine learning-based tumor probability maps computed from pre-biopsy mpMR images for predicting and visualizing potential biopsy targets representing clinically significant cancer foci. The median [range] AUC, sensitivity and specificity of the classifier were 0.87 [0.82–0.92], 0.77 [0.71-0.83] and 0.82 [0.76-0.86], respectively. This approach has a potential to reduce the number of biopsy cores, and thus the risk of post-biopsy infection/complications.

Background

mpMRI constitutes an integral part of targeted prostate biopsy sampling, either in-bore MRI-guided or ultrasound fusion-guided, for histopathological assessment and prostate cancer diagnosis 1,2. This technique has resulted in improved cancer detection rates 3,4. However, it still depends on qualitative evaluation and subjective judgement of experienced radiologists to determine lesions to be biopsied. Methods that can precisely and objectively detect clinically significant lesions could assist the radiologist in efficiently pinpointing the optimal biopsy target, thereby potentially reducing the number unnecessary biopsies and associated side effects. The aim of this study was to develop a machine-learning method for predicting and visualizing potential biopsy targets representing clinically significant cancer foci.

Materials and Methods

The dataset for this study was obtained from the training dataset from the ProstateX challenge 5. Pre-biopsy mpMRI (3T Magnetom Trio or Skyra; Siemens Medical Solutions, Erlangen, Germany) was performed in 188 men as part of routine clinical diagnosis of prostate cancer. The images were examined by experienced radiologists who segmented the whole prostate volume, and annotated suspicious cancerous volumes of interest (VOIs) on the T2-weighted images. Targeted biopsy cores were sampled from these volumes to ascertain the presence of clinically significant (Grade Group ≥ 2) 6 prostate cancer, leading to 64 true positive (TP) and 234 false positive (FP) radiological findings. Quantitative image features comprising standardized (Gaussian normalization) T2-weighted image intensities, apparent diffusion coefficient (ADC), calculated high b-value image intensities, and volume transfer constant (Ktrans) were extracted from the voxels in the TP and FP VOIs and from randomly sampled healthy voxels (≈500 per patient) after co-registration 7 into T2-weighted image space. A partial least squares-discriminant analysis (PLS-DA) classifier was trained (80% of data) and tested (20% of dataset) on these features to predict the probability of prostate voxels being clinically significant disease or not. The training employed 5-fold cross-validation for Latent-variable optimization. For robustness, the experiment was repeated 10 times with different training and test sets. The tumor probability maps were back projected into the T2-weighted image space to visualize the correspondence with the annotations. Mann-Whitney U test was used to evaluate the differences in mean probabilities of the TP, FP, and healthy VOIs. Area under receiver-operating characteristic curves (AUCs) were estimated as a measure of predictive performance of the classifier.

Results

A total of 298 suspicious volumes were annotated and targeted for biopsy; 133 in the transition zone and 165 in the peripheral zone. Clinically significant cancer was found in 58 men, constituting 64 (≈21%) of the annotations (36 transition zone and 28 peripheral zone cancers). The mean tumor probability was significantly higher (p < 0.05) for TP VOIs compared to FP and healthy VOIs (Figure 1a), resulting in better visualization of candidate cancerous lesion targets (Figure 2). The median [range] AUC, sensitivity and specificity of the classifier were 0.87 [0.82–0.92], 0.77 [0.71-0.83] and 0.82 [0.76-0.86], respectively (Figure 1b & c).

Discussion

Despite the improvement offered by the integration mpMRI, the number of negative biopsy cores remains high, which increases the risk of post-biopsy infection and complications due to multidrug-resistance 8. Also, since biopsy sampling is limited to small portions of what is commonly a spatially heterogeneous and multifocal disease, it is crucial that biopsies are targeted to the most aggressive cancer to avoid underestimation of the disease extent. We have established tumor probability map computed from mpMR images that are capable of predicting clinically significant cancer spots with 0.77 [0.71-0.83] sensitivity and 0.81 [0.76-0.86] specificity. This approach could offer two practical advantages to complement the radiological reading. First, it combines mpMR images into a single feature map, which reduces the number different imaging modalities to be evaluated and thus workload. Secondly, the approach is quantitative, objective and could be used to rule out candidate lesions for biopsy, thereby reducing number of biopsies and side effects. In future we seek to investigate clinical the feasibility and efficacy of the method.

Conclusion

Machine learning-based tumor probability maps are capable of complementing radiological reading in detection of cancer foci to guide targeted biopsy sampling. This can help reduce unnecessary number of biopsy cores, and thus the risk of post-biopsy infection and complications.

Acknowledgements

Data used in this research were obtained from The Cancer Imaging Archive (TCIA) sponsored by the SPIE, NCI/NIH, AAPM, and Radboud University

References

1. Pinto PA, Chung PH, Rastinehad AR, et al. J Urol. 2011;186(4):1281–1285.

2. Hambrock T, Somford DM, Hoeks C, et al. J Urol. 2010;183(2):520–528.

3. Pokorny MR, De Rooij M, Duncan E, et al. Eur Urol. 2014;66(1):22–29.

4. Thompson JE, Moses D, Shnier R, et al. J Urol. 2014;192(1):67–74.

5. Litjens G, Debats O, Barentsz J, et al. ProstateX Challenge data. Cancer Imaging Arch.

6. Epstein JI, Zelefsky MJ, Sjoberg DD, et al. Eur Urol. 2015;69(3):428–435.

7. Klein S, Staring M, Murphy K, et al. IEEE Trans Med Imaging. 2010;29(1):196–205.

8. Liss MA, Taylor SA, Batura D, et al. J Urol. 2014;192(6):1673–1678.

Figures

Figure 1: a) Box plots comparing the distribution of mean tumor probabilities across true positive (TP), false positive (FP) and healthy volumes. b) ROC curves showing performance of the PLSDA classifier for the 10 random runs. For each iteration, the dataset was randomly partitioned to give different training and test sets. c) Box plots showing the area under curves (AUC), sensitivity and specificity ranges of the classifier for each iteration. The threshold for true positive was obtained as optimal threshold from the ROC curve.

Figure 2: Examples of tumor probability maps of the prostate predicted from the PLS-DA classifier, and back projected into their corresponding T2-weighted image spaces. The red outline indicates regions that were marked by the radiologist as potential clinically significant cancers and were confirmed by biopsy (true positive), while the blue indicates regions rebutted by biopsy (false positive). c Indicates a patient with biopsy Grade Group 2.

Proc. Intl. Soc. Mag. Reson. Med. 27 (2019)
1846