Textured-Based Deep Learning in Prostate Cancer Classification with 3T Multiparametric MRI: Comparison with PI-RADS-Based Classification

Abstract

The current standardized scheme for interpreting MRI requires a high level of expertise and exhibits a significant degree of inter-reader and intra-reader variability. An automated prostate cancer (PCa) classification can improve the ability of MRI to assess the spectrum of PCa. The purpose of the study was to evaluate the performance of a texture-based deep learning model (Textured-DL) for differentiating between clinically significant PCa (csPCa) and non-csPCa and to compare the Textured-DL with Prostate Imaging Reporting and Data System (PI-RADS)-based classification (PI-RADS-CLA), where a threshold of PI-RADS ≥ 4, representing highly suspicious lesions for csPCa, was applied. The study cohort included 402 patients (60% (n = 239) of patients for training, 10% (n = 42) for validation, and 30% (n = 121) for testing) with 3T multiparametric MRI matched with whole-mount histopathology after radical prostatectomy. For a given suspicious prostate lesion, the volumetric patches of T2-Weighted MRI and apparent diffusion coefficient images were cropped and used as the input to Textured-DL, consisting of a 3D gray-level co-occurrence matrix extractor and a CNN. PI-RADS-CLA by an expert reader served as a baseline to compare classification performance with Textured-DL in differentiating csPCa from non-csPCa. Sensitivity and specificity comparisons were performed using Mcnemar's test. Bootstrapping with 1000 samples was performed to estimate the 95% confidence interval (CI) for AUC. CIs of sensitivity and specificity were calculated by the Wald method. The Textured-DL model achieved an AUC of 0.85 (CI [0.79, 0.91]), which was significantly higher than the PI-RADS-CLA (AUC of 0.73 (CI [0.65, 0.80]); p < 0.05) for PCa classification, and the specificity was significantly different between Textured-DL and PI-RADS-CLA (0.70 (CI [0.59, 0.82]) vs. 0.47 (CI [0.35, 0.59]); p < 0.05). In sub-analyses, Textured-DL demonstrated significantly higher specificities in the peripheral zone (PZ) and solitary tumor lesions compared to the PI-RADS-CLA (0.78 (CI [0.66, 0.90]) vs. 0.42 (CI [0.28, 0.57]); 0.75 (CI [0.54, 0.96]) vs. 0.38 [0.14, 0.61]; all p values < 0.05). Moreover, Textured-DL demonstrated a high negative predictive value of 92% while maintaining a high positive predictive value of 58% among the lesions with a PI-RADS score of 3. In conclusion, the Textured-DL model was superior to the PI-RADS-CLA in the classification of PCa. In addition, Textured-DL demonstrated superior performance in the specificities for the peripheral zone and solitary tumors compared with PI-RADS-based risk assessment.

Keywords: PI-RADS; convolutional neural network; deep learning; prostate cancer classification; texture analysis.

Figure 1

The overall workflow of the proposed Textured-DL model for the PCa classification. Suspicious prostate lesions were first detected and scored by the PI-RADS, followed by contouring. Then, 3D volumetric patches of the prostate lesion were cropped from the T2W and ADC images, and GLCM were extracted from two patches. Next, the two GLCMs were concatenated and fed into CNN to generate the probability of csPCa. ROC curve, AUC, sensitivity, and specificity were adopted to evaluate and compare the performance of PCa classification by the PI-RADS-CLA and Textured-DL, which was confirmed by the histopathological findings. In the GLCM extractor, each cubic box represents a voxel. The distance between adjacent voxels was enlarged to see the directions between voxels clearly. (i, j, k) is a voxel coordinate, and (i, j + 1, k) is an immediate neighboring voxel (INV) coordinate around the voxel.

Figure 2

Two examples of prostate lesion classifications are shown in rows (a,b), respectively. From left to right, axial T2W, axial ADC, GLCMs derived from the prostate lesion volumetric patch and matched WMHP are shown in each row. (a) Imaging for a 56-year-old man with a PSA of 12.2 ng/mL. A lesion (blue rectangular box pointed by a red arrow) with a PI-RADS score of 4 and GS 3 + 3 was shown on both the axial T2W and ADC images. The proposed Textured-DL predicted this lesion as a non-csPCa. (b) Imaging for a 72-year-old man with a PSA of 8.8 ng/mL. A lesion (blue rectangular box pointed by a red arrow) with a PI-RADS score of 3 and GS 4 + 3 was shown on both the axial T2W and ADC images. The Textured-DL predicted this lesion as a csPCa.

Figure 3

Comparisons of ROC (a), sensitivity (b), and specificity (b) between PI-RADS-CLA, Textured-RF, Imaged-RF, DTL, DCNN, and Textured-DL on the classification of PCa in the overall tumor lesions. CIs are provided in the square brackets. p values for the comparison between the baseline methods and the Textured-DL are also shown.

Figure 4

Comparisons of ROC, AUC, sensitivity, and specificity between the Textured-DL and baseline methods in the tumor lesions on the PZ (a1,a2) and TZ (b1,b2). CIs are provided in the square brackets. p values for the comparison between the baseline methods and the Textured-DL are also shown.

Figure 5

Comparisons of ROC, AUC, sensitivity, and specificity between the Textured-DL and baseline methods in the solitary (a1,a2) and multi-focal tumors (b1,b2). CIs were provided in the square brackets. p values for the comparison between the baseline methods and the Textured-DL were also provided.