Accurate assessment of CD is pivotal for patient management [23,24,25]. MRE provides non-invasive insights into both structural and functional aspects without ionizing radiation. Objective endpoints for CD management have been evaluated and established in the form of MRE indices [10, 11, 26]. However, prior research raises concern about MRE analysis being prone to human error, stemming from the time and attention required in radiologists’ interpretations of the entire bowel [27].

The complex nature of CD can benefit from more reliable assessment of disease activity, distribution, and treatment response. This is especially true for clinical trials, where precise quantification of disease is essential. Interestingly, our review demonstrates varied clinical tasks of deep learning applied to CD in MRE. These range from improving image quality, segmenting disease to quantify burden, and 3D reconstruction for surgery planning.

The use of deep learning models presents significant potential advancements in accuracy with these algorithms applied to MRE image analysis for CD [12, 13, 18]. The two included studies in this review evaluated image quality using deep learning applications: Son et al. attempted a DLR technique on 35 patients with known IBD, while Lian et al. employed CNN optimization on 392 patients with suspected IBD. Both studies reported a significantly increased quantitative signal-to-noise ratio. Son et al. demonstrated a qualitative improvement in image quality, including contrast, sharpness, motion artifacts, and blurring in the coronal and axial images; however, DLR image sets appeared more synthetic [18]. Lian et al. demonstrated a lower high-frequency error norm with the optimized CNN algorithm compared to other algorithms and reported a sensitivity of 95.2% and specificity of 46.7% for epidemic IBD, indicating significantly increased image quality with the use of CNN optimization compared to baseline [19].

Lamash et al. evaluated the effectiveness of CNN-based segmentation in 23 pediatric patients [20]. Aside from showcasing identification of active from non-active bowel segments, similar to conventional methods, deep learning models were able to accurately, and automatically, identify bowel strictures. Stricture length measurements between two senior radiologists were also reported with greater agreement.

CD, among other gastrointestinal diseases, is associated with deviations in small bowel motility [28,29,30]. A novel study has emerged for the quantitative assessment of 3D cine-MRI of small intestinal motility in both healthy patients and those with severe IBD [21]. By employing a combination of stochastic tracking and CNN-based orientation classifiers, Van Harten et al. successfully differentiated between motile and non-motile bowel segments. Their findings indicated that such models may surpass the clinically recommended model in assessing ileal CD activity. This highlights the potential for precise, automated, and non-invasive monitoring of intestinal inflammation in CD patients [27].

McFarlane et al. developed a CNN-based algorithm to assist in 3D image reconstruction, enhancing surgical planning for four patients with complex perianal fistulas in Crohn’s disease [22]. The reconstruction models provided a comprehensive representation of the perianal disease, aiding in the identification of internal fistula orifices, seton placements, and fistula tracts. While not providing extra information beyond what is acquired through MRI, 3D image reconstruction offers a more realistic depiction of the anatomy. This study demonstrates the potential of CNN-based diagnostic tools to assist surgeons improving surgical outcomes and reducing the number of surgical procedures needed for patients.

The heterogeneous focus of the included studies highlights the broad potential of deep learning in CD assessment via MRE. However, this diversity also complicates direct comparison of outcomes. Despite these challenges, the variety underscores the versatility of deep learning technologies in improving diagnosis, monitoring, and treatment planning for CD. To harness this potential fully, future research must aim for more standardized study designs. Such consistency will facilitate clearer comparisons and enable validation of deep learning applications across clinical settings. This approach will not only streamline research efforts but also accelerate the integration of these technologies into clinical practice, offering new avenues for patient care in CD management.

The heterogeneity observed in the studies under review reflects the multifaceted applications of deep learning across the spectrum of CD management via MRE, highlighting its potential to revolutionize the patient journey in radiology. From improving images through disease burden quantification to treatment planning, deep learning could potentially enhance each step, offering a more nuanced, precise, and patient-centered approach.

Despite the potential on display, research around deep learning is still in its early days, with few studies tackling these varied applications. Analyzing this evolving field of research is crucial. As deep learning and MRE technology advance, they promise to offer deeper insights into CD, enhancing diagnosis, monitoring, and surgical planning. Continuously reviewing these developments is essential for harnessing their full potential to improve patient care.

This systematic review, while providing insights into the emerging role of deep learning in enhancing MRE for CD assessment, has several limitations that warrant discussion:

Heterogeneity of included studies: The diversity in the objectives, methodologies, and outcomes of the included studies presents a significant challenge in synthesizing findings. This heterogeneity stems from varied focuses, such as image quality improvement, disease burden quantification, and 3D reconstruction for surgical planning. While this breadth highlights the potential of deep learning across different aspects of CD imaging, it complicates direct comparisons and synthesis of results and prevents performing a meta-analysis. Our rationale for including these diverse studies was to capture this research field’s current directions, acknowledging that deep learning in CD MRE is rapidly evolving with research in its early stages.

Methodological diversity: The included studies exhibit a range of study designs, patient populations, and deep learning techniques, which may influence the generalizability and comparability of findings. The predominance of retrospective studies and the absence of external validation in the analyzed research further contribute to potential biases and limit the strength of our conclusions.

Quality and relevance assessment: Our systematic review adhered to established guidelines and employed the QUADAS-2 tool for assessing the risk of bias. However, the subjective nature of some assessment criteria and the limited number of studies meeting our inclusion criteria may have impacted the robustness of our evaluation.

Addressing the limitations identified in this review requires an effort toward conducting prospective, multicenter studies with larger sample sizes, and external validation. Standardizing outcome measures and employing consistent deep learning methodologies will facilitate more direct comparisons of findings. Future reviews should also focus on overcoming the challenges of study heterogeneity by exploring specific tasks and specific algorithms of deep learning in CD MRE.