ERS3054. Deep Learning-Based MR Image Quality Assessment and its Impact on Prostate Cancer Detection Rates
Authors* Denotes Presenting Author
Yue Lin *;
National Cancer Institute, National Institutes of Health
Mason Belue;
National Cancer Institute, National Institutes of Health
Enis Yilmaz;
National Cancer Institute, National Institutes of Health
Stephanie Harmon;
National Cancer Institute, National Institutes of Health
Peter Choyke;
National Cancer Institute, National Institutes of Health
Baris Turkbey;
National Cancer Institute, National Institutes of Health
Objective:
A key element in detecting and ruling out clinically significant (CS) prostate cancer is the quality of the MR images. Prebiopsy T2-weighted (T2W) image quality is particularly important as it is used for spatially registering lesion contours to ultrasound (US) for MRI/US fusion-guided targeted biopsy (TBx). Artificial intelligence (AI) has the potential to assist radiologists in objectively determining MR image quality. The purpose of this study is to use a deep learning AI algorithm to classify prostate MR images in terms of quality and to examine the impact of image quality on prostate cancer detection.
Materials and Methods:
This retrospective study included patients who were imaged with multiparametric MRI and subsequently received combined TBx and 12-core US-guided systematic biopsy (SBx) for prostate cancer interrogation from April 2019 to September 2022. All scans were prospectively evaluated and intraprostatic lesions were scored using PI-RADSv2.1. For each patient, both T2W images and ADC maps were classified as diagnostic vs. non-diagnostic by a previously developed AI algorithm (trained from N=732 cases, achieving 83.9% accuracy for T2W images and 84.6% accuracy for ADC maps). Patient-based cancer detection rates (CDRs) were calculated for CS prostate cancer (Gleason Grade Group > 1) involvement. Chi-squared tests were performed to compare CS CDRs between diagnostic and non-diagnostic images in each PI-RADSv2.1 category.
Results:
A total of 658 patients (median age 67 [range 35-82] years; median PSA 6.8 [0.5-197.2] ng/mL) with 1,029 MRI visible lesions (66.5% tumor-positive) were evaluated. For T2W image quality analysis, 247 (37.5%) sequences were classified as non-diagnostic and 411 (62.5%) were categorized as diagnostic by the algorithm. For ADC maps, 222 (33.7%) were non-diagnostic while 436 (66.3%) were diagnostic. TBx with diagnostic T2W images resulted in significantly higher CS CDR than non-diagnostic images for PI-RADSv2.1 category 4 lesions (50.7% vs. 33.7%, p=0.013). However, when SBx results were also taken into consideration, CS CDRs were comparable between diagnostic and non-diagnostic T2W images (55.2% vs. 44.2%, p=0.11) on combined biopsies. No significant difference in CDRs were observed between diagnostic and non-diagnostic ADC maps across all PI-RADSv2.1 categories on both TBx and combined biopsies.
Conclusion:
This study employed a deep learning AI algorithm to classify image quality of prostate MRI. The results indicate that higher quality T2W images are associated with better TBx-based CS cancer detection performance for PI-RADSv2.1 category 4 lesions. In patients with low quality T2W images, combining TBx with SBx may be necessary to compensate for the non-optimal performance of TBx.
