ARRS 2022 Abstracts


ERS3344. Comparison of Quality and Lesion Detection of Deep Learning-Based T2-Weighted Reconstruction to Clinical T2-Weighted Images on Prostate MRI
Authors * Denotes Presenting Author
  1. Barun Bagga *; New York University School of Medicine
  2. Robert Petrocelli; New York University School of Medicine
  3. Paul Smereka; New York University School of Medicine
  4. Hersh Chandarana; New York University School of Medicine
  5. Angela Tong; New York University School of Medicine
The purpose of the study was to compare image quality and lesion detection of a prototype accelerated T2 weighted deep learning reconstruction (DL-T2) sequence compared to a clinical T2-weighted TSE (C-T2) on biparametric MRI of the prostate.

Materials and Methods:
DL-T2 (Siemens Healthineers, Erlangen, Germany) is an accelerated deep learning-based reconstruction, with k-space under-sampling via GRAPPA parallel imaging sampling, resulting in a reduction of average acquisition time from 3 min 22 seconds to 49 seconds. We retrospectively evaluated 80 consecutive patients (ages 47-84, PSA 0.4-25.4ng/mL) who underwent prostate MRI with C-T2 and DL-T2. All had a prostate biopsy within 6 months of MRI or stability of PSA or MRI for at least a year. C-T2 and DL-T2 images were anonymized and randomly presented to two blinded readers for image quality assessment on a Likert scale of 1(non-diagnostic)-4(excellent). Each reviewer evaluated bi-parametric MRI consisting of clinical diffusion-weighted images and either C-T2 or DL-T2 utilizing PI-RADS v2.1. Wilcoxon signed-rank test was used to compare the image quality scores. McNemar test and receiver operator characteristic (ROC) curve analysis were performed to compare lesion detection.

Biopsy results were available for 55/80 patients (18/55 = Gleason Grade Group 2 or higher). 25 patients had PI-RADS 1 or 2 results on prostate MRI with stable follow up PSA or MRI. The image quality results were as follows; Reader 1: overall quality (mean+/-SD: 3.88+/-0.40 (DL-T2); 3.73+/-0.53 (C-T2); p-value=0.03), clarity of capsule (mean+/-SD: 3.91+/-0.33 (DL-T2); 3.76+/-0.51 (C-T2); p-value=0.02), clarity of peripheral and transitional zone boundary (mean+/-SD: 3.91+/-0.33 (DL-T2); 3.76+/-0.51 (C-T2); p-value=0.02), and clarity of periurethral area (mean+/-SD: 3.85+/-0.39 (DL-T2); 3.74+/-0.52 (C-T2); p-value=0.09). Reader 2, overall quality (mean+/-SD: 3.31+/-0.74 (DL-T2); 3.33+/-0.82 (C-T2); p-value=0.97), clarity of capsule (mean+/-SD: 3.53+/-0.69 (DL-T2); 3.46+/-0.75 (C-T2); p-value=0.46), clarity of peripheral and transitional zone boundary (mean+/-SD: 3.19+/-0.84 (DL-T2); 3.28+/-0.83 (C-T2); p-value=0.39), and clarity of periurethral area (mean+/-SD: 3.34+/-0.75 (DL-T2); 3.36+/-0.77 (C-T2); p-value=0.81). Intra-class correlation coefficient for overall image quality was fair-to-good for DL-T2 (ICC: 0.57; range: 0.33-0.72), and was good for C-T2 (ICC: 0.72; range: 0.56 - 0.82). Frequency of lesion detection was not significantly different between DL-T2 + DWI (94 lesions) and C-T2 +DWI (91 lesions), p-value=0.77. We utilized PI-RADS 3 or greater as positive for prostate cancer. For reader 1, area under the curve (AUC) of ROC: 0.75 (clinical exam (CE)), 0.75 (deep learning exam (DLE)); Sensitivity (Sn): 0.61 (CE), 0.61 (DLE); Specificity (Sp): 0.74 (CE), 0.81 (DLE);. For reader 2, AUC of ROC: 0.65 (CE), 0.63 (DLE); Sn: 0.56 (CE), 0.61(DLE); Sp: 0.61 (CE), 0.50 (DLE).

Novel accelerated DL-T2 has the potential to reduce scan time for prostate MRI and shows comparable image quality and diagnostic performance to clinical T2-weighted imaging.