Feasibility of AI-assisted compressed sensing protocols in knee MR imaging: a prospective multi-reader study

Magnetic resonance imaging (MRI) offers outstanding advantages for structural abnormalities of the knee, which is currently one of the most common examinations [1, 2]. A major challenge of MRI acquisition is the prolonged scan duration compared to other modalities, limiting patients’ throughput and potential indications. Longer scan duration would increase the probability of motion artifact interference, especially in patients with knee pain, even resulting in scan failure. Thus, developing advanced accelerated scan protocols is worthy of clinical investigation [3‐5].

Artificial intelligence (AI)–based accelerated MRI is in its infancy but has made significant progress in recent years [6]. Although several advanced techniques have been proposed to reduce the image acquisition time, parallel imaging (PI) is proven to reduce the acquisition time while maintaining diagnostic performance and is commercially available for routine clinical use [7, 8]. Increments in the acceleration factor of the conventional acceleration methods may cause reductions in the signal-to-noise ratio (SNR) and increased image artifacts, limiting the feasibility of further accelerated scanning [9, 10]. The recent development of AI-based reconstruction methods pre-learns image structure information from numerous fully sampled MR images via convolutional neural networks and applies the pre-trained knowledge to undersampled image reconstruction, allowing high acceleration levels while maintaining high image quality and potentially yielding a breakthrough in accelerated MRI scanning [6, 11‐13].

However, AI-based image reconstruction must be realistic to ensure correct diagnosis in clinical practice. Although some previous studies have investigated the application of AI-based reconstruction, including single sequence studies [14] and retrospective undersampled simulations with accelerated acquisition [15], the main limitation of studies to date is the absence of diagnostic information content [16‐18] and surgical ground truth, which limits the support for clinical application [19, 20]. Thus, there are still challenges for AI-based acceleration approaches to enter clinical practice. Further studies are needed to explore the reliability of AI-based acceleration methods in clinical practice.

This study aims at evaluating an AI-assisted compressed sensing (ACS) acceleration technique in the clinical practice of knee imaging, to determine whether ACS protocols can yield accelerated knee MRI scans without compromising the diagnostic value of images and can present advantages in improving the efficiency and accessibility of MRI through improved patient comfort and fewer motion artifacts. We also evaluated a 2.0-min accelerated protocol, which may hold promise for rapid knee scans of patients with special clinical needs.

Our study showed that artificial intelligence–assisted compressed sensing (ACS) can be used for reconstruction in knee MRI and performed as well with conventional 2-fold parallel imaging (PI) to detect internal knee derangement while reducing the acquisition time by even more than 70%. Moreover, ACS protocols showed good-to-excellent inter-reader and inter-method agreement with an improved signal-to-noise ratio (SNR) and reduced edge rise distance (ERD).

MRI is one of the most dynamic and safe imaging techniques in clinical practice today and offers an unparalleled soft tissue resolution. However, MRI is an inherently slow imaging modality, limited by traditional Nyquist sampling requirements, and prone to motion artifacts. With the increasing demand in clinical practice, acceleration of image data acquisition has been a focus of research, yielding several undersampling reconstruction techniques such as partial Fourier (PF), PI, simultaneous multislice (SMS) imaging, compressed sensing (CS), and AI-based reconstruction algorithms. Despite its wide clinical usage and integration into commercial MR systems, PI with multiple receiver coils typically requires limiting the acceleration factor to 2 while maintaining spatial resolution [9]. Similarly, the acceleration factor in SMS is generally limited to 4 due to slice crosstalk artifacts and increased energy deposition into the tissue through multislice excitation.

In contrast, undersampled reconstruction techniques such as CS can accelerate images beyond a factor of 4. The assumption underlying CS is that the k-space data are randomly undersampled. CS-based reconstruction techniques employ pseudorandom trajectories along with one or several sparse transforms such as finite differences or wavelet operators [26]. After receiving Food and Drug Administration approval in 2017, CS-based reconstruction is being gradually adopted into clinical practice. Convolutional neural network–based ACS acceleration technology will likely be the next entry point for accelerated imaging [13]. However, incorporating such protocols into clinical practice will require ongoing and extensive validation of deep learning models and a thorough understanding of their generalizability and limitations.

Musculoskeletal imaging is highly demanding regarding resolution, clarity, and image quality. However, patients with pain or a limited range of motion may experience difficulty in maintaining stable imaging positions. In our study, ACS technology showed clinical utility in accelerated knee MRI scan acquisition, which has been previously assessed in abdomen imaging of the liver [27]. The artificial intelligence–based filtering and interpolation image reconstruction techniques are based on coding and decoding convolutional neural networks to achieve high-resolution imaging without loss of time or image SNR by deep learning to remove noise for better image quality [12]. Our study included three accelerated scanning protocols based on two different technologies: the PI 8-min protocol and the ACS 3.5- and 2.0-min protocols. Image evaluation was performed by assessing image quality and diagnostic efficacy. The comparison of quantitative assessments and task-based (diagnostic) performance metrics showed that ACS-accelerated MRI scans are equal to or close to conventional PI protocols in diagnosing most structural knee disorders with reduced noise and significantly improved image sharpness and SNR. This study included a diverse range of patients with acute trauma, chronic injury, cruciate ligament reconstruction, and post-arthroscopic surgery, enhancing the representativeness of its findings for clinical practice.

Although AI reconstruction of MR images has been investigated previously, those studies compared sequence equivalence on single sagittal sequences of the knee [14] or lacked the diagnostic accuracy of the protocols [15, 19, 20]. To the best of our knowledge, this is the first study to compare ACS-accelerated protocols with a standard clinical MRI protocol in all four sequences (one T1-weighted FSE and three proton density–weighted FSE sequences with fat saturation) with the diagnostic accuracy of the protocols compared. Moreover, according to the indications of knee MRI application in actual clinical practice, the types of participants included in this study were more diverse than in previous studies. Participants after cruciate ligament surgery were included. Interobserver agreement was good for all three protocols, suggesting that ACS-accelerated MRI did not introduce more artifacts interfering with the observation of knees with certain metal implants and can be used for postoperative examinations. In addition, the absence of radiation makes knee MRI especially applicable in children with joint discomfort; thus, pediatric participants were included in this study for the examination.

Although the findings demonstrated the feasibility of the proposed ACS method, this study had some limitations. First, all images were obtained from a 3.0-T MRI scanner from a single vendor, and no validation was performed on other machines. Further studies at multiple vendors and other scanning sites are warranted to assess the scalability of AI-assisted acceleration technology fully. Second, only a small percentage of participants underwent arthroscopy and had pathological findings for diagnostic performance analysis, potentially diminishing the persuasiveness of the results. However, in clinical practice, not all patients with the detected disease require arthroscopy, and interobserver and inter-method agreements may address this concern. Third, the low incidence of some intra-articular knee lesions in our study cohort may have limited statistical efficacy. Necessary future steps include correlating readers’ assessments of pathological findings compared against ground truths (arthroscopy) and establishing robustness across the various scanners and vendors. Thus, additional research with a larger pool of patients from internal and multi-center studies is necessary to verify these findings. Further studies would also include validating the ACS acceleration technique on quantitative MRI, such as T1/T2 mappings and 3D sequences.

In conclusion, comparing the artificial intelligence–assisted compressed sensing (ACS) technique with parallel imaging (conventional acceleration factor of 2.0) illustrated that ACS acceleration methods using novel reconstruction algorithms can reduce scanning times and maintain high image quality and equivalent diagnostic performance.