The added value of AI-based computer-aided diagnosis in classification of cancer at prostate MRI.

OBJECTIVES

To develop an artificial intelligence (AI) model for prostate segmentation and prostate cancer (PCa) detection, and explore the added value of AI-based computer-aided diagnosis (CAD) compared to conventional PI-RADS assessment.

METHODS

A retrospective study was performed on multi-centers and included patients who underwent prostate biopsies and multiparametric MRI. A convolutional-neural-network-based AI model was trained and validated; the reliability of different CAD methods (concurrent read and AI-first read) were tested in an internal/external cohort. The diagnostic performance, consistency and efficiency of radiologists and AI-based CAD were compared.

RESULTS

The training/validation/internal test sets included 650 (400/100/150) cases from one center; the external test included 100 cases (25/25/50) from three centers. For diagnosis accuracy, AI-based CAD methods showed no significant differences and were equivalent to the radiologists in the internal test (127/150 vs. 130/150 vs. 125/150 for reader 1; 127/150 vs.132/150 vs. 131/150 for reader 2; all p > 0.05), whereas in the external test, concurrent-read methods were superior/equal to AI-first read (87/100 vs. 71/100, p < 0.001, for reader 2; 79/100 vs. 69/100, p = 0.076, for reader 1) and better than/equal to radiologists (79/100 vs. 72/100, p = 0.039, for reader 1; 87/100 vs. 86/100, p = 1.000, for reader 2). Moreover, AI-first read/concurrent read improved consistency in both internal test (κ = 1.000, 0.830) and external test (κ = 0.958, 0.713) compared to radiologists (κ = 0.747, 0.600); AI-first read method (8.54 s/7.66 s) was faster than readers (92.72 s/89.54 s) and concurrent-read method (29.15 s/28.92 s), respectively.

CONCLUSION

AI-based CAD could improve the consistency and efficiency for accurate diagnosis; the concurrent-read method could enhance the diagnostic capabilities of an inexperienced radiologist in unfamiliar situations.

KEY POINTS

• For prostate cancer segmentation, the performance of multi-small Vnet displays optimal compared to small Vnet and Vnet (DSC vs. DSC, p = 0.021; DSC vs. DSC, p < 0.001). • For prostate gland segmentation, the mean/median DSCs for fine and coarse segmentation were 0.91/0.91 and 0.88/0.89, respectively. Fine segmentation displays superior performance compared to coarse (DSC vs. DSC, p < 0.001). • For PCa diagnosis, AI-based CAD methods improve consistency in internal (κ = 1.000; 0.830) and external (κ = 0.958; 0.713) tests compared to radiologists (κ = 0.747; 0.600); the AI-first read (8.54 s/7.66 s) was faster than the readers (92.72 s/89.54 s) and the concurrent-read method (29.15 s/28.92 s).