Deep Learning-Enhanced T1-Weighted Imaging for Breast MRI at 1.5T

Abstract

Background/Objectives: Assessment of a novel deep-learning (DL)-based T1w volumetric interpolated breath-hold (VIBE_DL) sequence in breast MRI in comparison with standard VIBE (VIBE_Std) for image quality evaluation. Methods: Prospective study of 52 breast cancer patients examined at 1.5T breast MRI with T1w VIBE_Std and T1 VIBE_DL sequence. T1w VIBE_DL was integrated as an additional early non-contrast and a delayed post-contrast scan. Two radiologists independently scored T1w VIBE _Std/DL sequences both pre- and post-contrast and their calculated subtractions (SUBs) for image quality, sharpness, (motion)-artifacts, perceived signal-to-noise and diagnostic confidence with a Likert-scale from 1: Non-diagnostic to 5: Excellent. Lesion diameter was evaluated on the SUB for T1w VIBE_Std/_DL. All lesions were visually evaluated in T1w VIBE_Std/_DL pre- and post-contrast and their subtractions. Statistics included correlation analyses and paired t-tests. Results: Significantly higher Likert scale values were detected in the pre-contrast T1w VIBE_DL compared to the T1w VIBE_Std for image quality (each p < 0.001), image sharpness (p < 0.001), SNR (p < 0.001), and diagnostic confidence (p < 0.010). Significantly higher values for image quality (p < 0.001 in each case), image sharpness (p < 0.001), SNR (p < 0.001), and artifacts (p < 0.001) were detected in the post-contrast T1w VIBE_DL and in the SUB. SUB_DL provided superior diagnostic certainty compared to SUB_Std in one reader (p = 0.083 or p = 0.004). Conclusions: Deep learning-enhanced T1w VIBE_DL at 1.5T breast MRI offers superior image quality compared to T1w VIBE_Std.

Keywords: MRI; breast cancer; deep learning; diagnostic imaging.

Figure 1

Overview of the entire study cohort examined via 1.5T breast MRI with T1 VIBE Standard (Std) and T1 VIBE deep learning (DL). No special type (NST); ductal carcinoma in situ (DCIS).

Figure 2

Flowchart of the reconstruction process using undersampled k-space data as well as precalculated coil sensitivity maps as input, performing the image generation through six iterations that alternate between a parallel imaging model and a network-based image enhancement, and finally interpolating the image using a deep learning-based super-resolution tailored to the selected partial Fourier sampling in the acquisition. Furthermore, the conventional 3D U-net architecture used in each iteration is detailed with the employed hyperparameters.

Figure 3

Patient with histologically proven no-special type (NST) breast carcinoma G2 on the left side, with a lesion diameter of 45mm (pT2). Compared to T1 VIBE_Std pre-contrast (a), post-contrast (c), and the last SUB (e), the T1 VIBE_DL pre-contrast (b), post-contrast (d), and the SUB (f) display the tumor with higher sharpness, without affecting diagnostic confidence. Based on the evaluation of the latest SUB, there are relevant motion artifacts in T1 VIBE SUB_Std and _DL, based on the study design.

Figure 4

Patient with histologically proven invasive lobular breast carcinoma (ILC) G2 on the left side. Tumor margins are sharper in T1 VIBE_DL (b) compared to T1 VIBE_Std (a) in pre-contrast images. No tumor assessment is possible in non-diagnostic post-contrast images in T1 VIBE_Std and _DL (c,d). The tumor is displayed sharply in T1 VIBE SUB_DL compared to T1 VIBE SUB_Std (e,f), while more artefacts outside the area of interest were detected in T1 VIBE_DL, e.g., in the mediastinum (g,h).