Deep Learning HASTE for Upper Abdominal MRI: Improved Image Quality, Speed, and Energy Efficiency in a Prospective Study

Abstract

This prospective study aimed to perform a qualitative and quantitative comparison of deep learning (DL) and conventional T2-weighted Half-Fourier Acquisition Single-shot Turbo spin Echo (HASTE) sequences for 3T MRI acquisition of the upper abdomen. From January 2024 to April 2024, 166 patients (60 ± 14 years) scheduled for MRI of the upper abdomen were prospectively enrolled. Each patient underwent two MRI examinations: one using a conventional T2-weighted HASTE sequence, followed by a fast T2-weighted HASTE sequence reconstructed with DL. Image quality, anatomical structure visualization, and diagnostic performance were independently assessed by three readers using a 5-point Likert scale. Quantitative analysis included measurements of signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) for both sequences. Additionally, radiomic features were extracted and analyzed for significant variations. Interreader agreement was evaluated using Fleiss' Kappa. The DL HASTE sequence showed significantly superior overall image quality (p < 0.001), fewer artifacts (p < 0.001), and improved delineation of anatomical structures (p < 0.01) compared to the conventional T2-weighted HASTE sequence. DL sequences exhibited better SNR (p < 0.001), whereas CNR values did not show a difference between the two acquisition types. Radiomics feature analysis unveiled significant differences in contrast and gray-level characteristics (p ≤ 0.001). DL HASTE demonstrated a significant time reduction of 62.5% together with significant energy cost savings of 0.34 kW per scan compared to the conventional sequence acquisition. The DL HASTE sequence enhanced image quality and diagnostic confidence while minimizing artifacts, time, and energy costs, enabling a more accurate detection of pathologies than the conventional T2-weighted product solution with potential clinical impact.

Keywords: deep learning; diagnostic imaging; magnetic resonance imaging; neural network.

FIGURE 1

MRI of a 46‐year‐old female with a palpable mass of the upper right abdomen showing inhomogeneous liver parenchyma with multiple cystic lesions in segments 4a/b and 5–7 and multiple smallnodular foci. (a) Axial, noncontrast conventional HASTE sequence; (b) axial, noncontrast DL‐HASTE sequence; (c) contrast‐enhanced T1‐sequence (venous phase); (d) coronal contrast‐enhanced T1 (venous phase); (e) diffusion map (TRACE); and (f) diffusion map (ADC).

FIGURE 2

MRI of a 56‐year‐old female with suspected Morbus Crohn of Ileum. (a) Axial T2‐HASTE sequence, (b) axial DL‐HASTE sequence, (c) axial T1 sequence, (d) coronal T1 sequence, (e) diffusion map (Trace), and (f) diffusion map (ADC).

FIGURE 3

Boxplot of SNR and CNR metrics. This boxplot displays the distribution of SNR (A) and CNR (B) metrics of the standard T2‐weighted HASTE sequence and the DL‐T2‐weighted HASTE sequence. Abbreviations: CNR, contrast‐to‐noise ratio; SNR, signal‐to‐noise ratio.