AI algorithmically-enhanced motion suppression simulating an osteochondral defect in a young child

Abstract

Artificial intelligence (AI) in radiology has rapidly increased in our field and stands to allow more accurate diagnosis, quicker interpretations, easier workflows, and improved image quality. However, with superior image quality produced with the help of AI algorithms, one could begin to discount or even eliminate the review of nonalgorithmic enhanced images. At least currently, these images remain important. This case report demonstrates a unique anomaly simulating disease resulting from AI-enhanced motion suppression. On the original images, patient motion and an atypical linear motion artifact is obvious. However, the images reproduced using our AI motion artifact suppression algorithm suppressed nearly all (but not all) of the motion artifact resulting in what appeared to be an osteochondral lesion in a child's knee. This case illustrates the necessity for the interpreting radiologist to review both original acquisitions as well as AI-enhanced images, at least for the time being.

Introduction

In the last decade, radiology has seen accelerated growth in the advent and deployment of artificial intelligence (AI) solutions for both interpretative and noninterpretive tasks [1], [2], [3]. On the whole, these technologies are likely a net positive for our field. Noninterpretative uses of AI in MRI include algorithms helping in noise reduction, decreasing MRI scan times, and motion reduction, among other uses [1]. As these algorithms improve and become widely available, one can see a time when the “original” images are devalued and, eventually, not even transmitted to the radiologist for review during study interpretation. A contemporary corollary to this future state is tomograms and CT. In other words, some institutions do not routinely send topograms along with other series to the interpreting radiologist during study interpretation, a practice which can be problematic [4].

We present a case demonstrating the current utility of original acquisition images in identifying artifacts that were subsequently (but not completely) suppressed by our MRI motion-suppression AI-generated images. In this case, proper identification of the artifact was paramount in avoiding a misdiagnosis of a femoral osteochondral lesion in a child.

Case report

A 12-year-old female gymnast presented to her pediatrician with a 4-day history of left knee pain status-post hyperextension injury. Pain was sudden in onset and rated 7/10 at the time of evaluation. Physical exam was positive for tenderness on palpation along the medial collateral ligament distribution and medial knee joint line. Patient experienced pain without laxity during McMurray's and valgus stress maneuvers. The remainder of the physical examination was noncontributory. Initial radiographs were normal, and an MRI was ordered and performed on a 3-Tesla MRI scanner (SIGNA™ Premier, GE Healthcare, Chicago, IL).

Initial images, including AI algorithmically-enhanced images, demonstrated focal edema in the medial femoral condyle with focal clefts through the adjacent articular cartilage, which appeared thickened (Fig. 1A). At the time of image interpretation, the radiologist was first presented with the algorithmically-enhanced images, although the nonalgorithmically enhanced images were included at the end of the study jacket (Fig. 1B). Additionally, repeat images were acquired during the study by the MRI technologist due to significant artifact present on the original T2-weighted images (Fig. 2). In our current workflows, MRI technologists are not presented with postprocessed algorithmically enhanced images during image acquisition. Thus, our technologists are expected to recognize suboptimal images and repeat series as necessary.

Fig. 1.

A 12 year-old female gymnast with acute onset knee pain following a hyperextension extension injury. (A and B) Sagittal T2 weight image with fat suppression with (A) and without (B) AI-enhanced motion suppression. (A) Image demonstrates focal clefting (arrowheads) through the lateral femoral condylar articular cartilage and subarticular cortex with adjacent bone marrow edema. In retrospect, there is a subtle vertical T2 hyperintense line (arrows) extending through the distal femoral epiphysis posteriorly and through the area of suspected cartilage clefting. (B) Without AI-algorithmic motion suppression, the linear T2-weighted artifact is shown to much greater advantage (arrows).

Fig. 2.

Repeat Sagittal T2-weighted image with fat-saturation demonstrates resolution of the motion artifact as well as the artifactual linear T2 hyperintense line. The cartilage and underlying subarticular cortex are intact.

As discussed above, our current PACs hanging protocols presents the interpreting radiologist with all AI-algorithmically enhanced protocols first. Due to space constraints, only those series are initially present. The radiologist must then load additional series within the study jacket individually for review. In this case, the interpreting radiologist reviewed and draft interpreted the patient as having an OCD lesion with cartilage defect in the lateral femoral condyle. As the radiologist worked through the rest of the patient's jacket, reviewing all available series, it became apparent that what he felt was an OCD, was in fact a combination of motion artifact and osseous contusion. On the algorithmically-enhanced, motion suppressed images, the linear motion artifact in the articular lateral femoral condylar bone (Fig. 1) was suppressed to the point where it was difficult to recognize initially. Only the cartilage cleft was initially apparent. Following subsequent review of additional images, the artifact was recognized, and the report was edited to reflect the osseous contusion, saving the potential for unnecessary arthroscopy. The patient did well with conservative management, with all symptoms resolving, and no additional symptoms reported.

Discussion

Currently, it would be hard to overestimate how important artificial intelligence has become in our society. The possibilities in radiology are equally as boundless. However, we currently are not at the point where AI in radiology can negate a vast majority of our duties. Rather it helps to augment our work.

In this case, without the astute technologist repeating the motion degraded sequences, the interpreting radiologist would be left in a conundrum because the AI-motion suppressed images are deceptive while the original images are suboptimal from motion. In essence, the examination could have been nondiagnostic in a busy practice where the time may not be available to repeat sequences. Alternatively, to the unwary radiologist, it is tempting to take AI-enhancement as ground truth and discount the original motion degraded images. Luckily in this case, the radiologist did not have to entertain this option because images without motion were obtained.

Following this case, our AI-algorithm has been further improved. To the authors’ knowledge, we have had no additional similar cases. Our radiologists are still presented with both the original and AI-enhanced series. Our technologists are still required to repeat suboptimal original images. Nevertheless, our ultimate goal is to both eliminate the need for repeat acquisitions as well as reduce the number of series interpreting radiologist must review. This aim will take time and further study, but we believe we will eventually accomplish this goal. Until then, we encourage all radiologists with AI-algorithmically enhanced images to continue to review original acquisitions until their algorithms are perfected.

Patient consent

The authors attest that the patient's parents provided written informed consent for the use of the patient's medical record and images for scientific purposes.