Improving the Quantitative Interpretation of Multi-parametric MRI (mp-MRI) in Prostate Cancer
- Author(s): Zhong, Xinran
- Advisor(s): Sung, Kyunghyun
- et al.
Prostate cancer (PCa) is the second leading cause of cancer-related death in men in the United States. The accurate diagnosis of PCa is crucial for proper treatment decision. Although biopsy is still the gold standard for diagnosis, it is limited to low sensitivity and invasiveness. On the other hand, as a non-invasive imaging tool, multi-parametric MRI (mp-MRI) has excellent potential in PCa diagnosis such as detection and stratification of aggressiveness. The mp-MRI includes both anatomical and functional information to be able to provide a comprehensive characterization of the tissue. However, diagnosis with mp-MRI is limited to inconsistent and qualitative interpretation. Clinically, the evaluation of mp-MRI is often through a standardized scoring system, PIRADS v2, which can lead to high inter- and intra-observer variability, and with a large amount of data for each case, the diagnosis process can be time-consuming. In order to get a more consistent quantitative evaluation, there are mainly two ways to utilize algorithms to help the diagnosis. The first one is creating quantitative biomarkers through mathematical models proposed based on assumption and understanding of physics and physiology, such as pharmacokinetic models for quantitative dynamic contrast-enhanced (DCE) MRI. The second one is using a machine learning technique to train a system with existing data to get diagnosis prediction on new data. The purpose of this work is to improve the quantitative interpretation of mp-MRI in PCa diagnosis regarding consistency and accuracy.
To evaluate existing B1+ estimation techniques to achieve a more consistent pre-contrast T1 estimation for quantitative DCE- MRI, 21 volunteers were prospectively recruited and scanned twice on two 3T MRI scanners, resulting in 84 variable flip angle (VFA) T1 exams. Two B1+ mapping techniques, including reference region variable flip angle (RR-VFA) and saturated turbo FLASH (satTFL), were used for B1+ correction, and T1 maps with and without B1+ correction were tested for the intra-scanner repeatability and inter-scanner reproducibility. Volumetric regions of interest were drawn on the transition zone, peripheral zone of the prostate and the obturator internus left and right muscles in the corresponding slices. The average T1 within each ROI for each scan was compared for both intra- and inter-scanner variability using concordance correlation coefficient and Bland-Altman plot.
To simplify B1+ compensation for quantitative DCE MRI in clinical and clinical research settings, an analytical B1+ correction method is proposed using a Taylor series approximation to the steady-state spoiled gradient echo signal equation. The proposed approach only requires B1+ maps and uncorrected pharmacokinetic (PK) parameters as the input, and was evaluated using a prostate digital reference object (DRO) and 82 in-vivo prostate DCE-MRI cases. The approximated analytical correction was compared with the ground truth PK parameters in simulation, and compared with the reference numerical correction in in-vivo experiments, using percentage error as the metric.
To develop a deep transfer learning (DTL) based model to distinguish indolent lesions from clinically significant PCa lesions using multiparametric MRI, 140 patients with 3T mp-MRI and whole-mount histopathology (WMHP) were included as the study cohort with IRB approval. The DTL based model was trained on 169 lesions in 110 arbitrarily selected patients and tested on the remaining 47 lesions in 30 patients. We compared the DTL based model with the same deep learning (DL) model architecture trained from scratch and the classification based on PIRADS v2 score with a threshold of 4 using accuracy, sensitivity, specificity, and area under curve (AUC). Bootstrapping with 2000 resamples was performed to estimate the 95% confidence interval (CI) for AUC.
Both RR-VFA-corrected T1 and satTFL-corrected T1 showed higher intra- and inter-scanner correlation (0.89/0.87 and 0.87/0.84 respectively) than VFA T1 (0.84 and 0.74). Bland-Altman plots showed that VFA T1 had a wider 95% limits of agreement and a larger range of T1 for each tissue compared to T1 with B1+ correction.
The prostate DRO results show that the proposed approach provides residual error less than 0.4% for both Ktrans and ve, compared to the ground truth. This noise-free residual error was smaller than the noise-induced error using the reference numerical correction, which had a minimum error of 2.1�4.3% with baseline SNR of 234.5. For the 82 in-vivo cases, percentage error compared to the reference numerical correction method had a mean of 0.1% (95% central range of [0.0%, 0.2%]) across the prostate volume.
After training on 169 lesions in 110 patients, the AUC of discriminating indolent from clinically significant PCa lesions of the DTL based model, DL model without transfer learning and PIRADS v2 score > 4 were 0.726 (CI [0.575, 0.876]), 0.687 (CI [0.532, 0.843]) and 0.711 (CI [0.575, 0.847]), respectively in the testing set.
The application of B1+ correction (both RR-VFA and satTFL) to VFA T1 results in more repeatable and reproducible T1 estimation than VFA T1. This can potentially provide improved quantification of the prostate DCE-MRI parameters.
The approximated analytical B1+ correction method provides comparable results with less than 0.3% error within 95% central range, compared to reference numerical B1+ correction. The proposed method is a practical solution for B1+ correction in prostate DCE-MRI due to its simple implementation.
The DTL based model achieved higher AUC compared to the DL model without transfer learning and PIRADS v2 score > 4 in discriminating clinically significant lesions in the testing set. The DeLong test indicated that the DTL based model achieved comparable AUC compared to the classification based on PIRADS v2 score (p = 0.89).