gms | German Medical Science

Objectives: An accurate diagnosis of muscle injuries as prerequisite of an efficient and successful treatment is one of the most difficult challenges in modern sports orthopaedics. Anamnesis, symptoms, palpation, functional testing and identifying the cause of muscle injuries are most crucial in order to initiate the correct treatment at the right time.

In 2013 the "Munich-Consensus Classification" was published in the British Journal of Sports Medicine, which classifies all relevant muscle injuries in professional sports into functional, structural and respective subgroups. MRI-based classification on the other hand only divides into four groups (1: small tear - 4: complete tear).

Functional injuries, especially neurogenic muscle hardening, are the most common muscle injuries. Based on the response of the athlete as well as MR scans, these seem very similar to structural injuries (muscle bundle tear) and this wrong diagnosis leads often falsely to week long breaks. Currently they can only be reliably identified through palpation of the muscle by a very experienced physician and a careful synopsis of clinical history and symptoms.

Our goal was to develop an artificial intelligence (AI) algorithm to precisely detect fine differences in MR images and distinguish between functional neurogenic muscle hardening and structural muscle bundle tear.

Methods: We collected 316 single MR slices (220 structural & 96 neurogenic injuries) from 31 patients without any specific localisation. We developed an algorithm adapted to the complexity of recognizing the two classes with a EfficientNet architecture. To tackle the limited and unbalanced amount of data, extensive data augmentation as well as an according loss weighting method was implemented.

We applied a data split of 70%, 20%, 10% for training, validation and testing, respectively. For statistical significance and robustness, data from a single patient was either part of training, validation or testing. During an additional 5-fold cross validation, random chosen test data for final evaluation remained untouched.

Results and Conclusion: As shown in Table 1 [Tab. 1], the algorithm achieved an accuracy of 85%/80% and a Dice Score of 83%/76% on validation/test data.

The GradCam in Figure 1 [Fig. 1] shows an activation map, which indicates the important regions for the AI predictions.

The results clearly state that there are differences between neurogenic muscle hardening and muscle bundle tear, such as multiple relevant regions in the case of neurogenic injuries, which can be detected by AI algorithms. Even with the currently very limited amount of data, the results present significant robustness and prediction accuracy.