gms | German Medical Science

Objectives: Ewing sarcomas (ES) are highly malignant tumors that are mainly found in children and adolescents. Early clinical diagnosis is the most important step but the difficulty of recognizing an ES is in its rarity and radiological peculiarity. Taking into consideration that the radiological signs can also be compatible with osteomyelitis, the chances of early diagnosis are even more difficult. Ewing sarcoma and osteomyelitis are two profoundly different diseases with very different treatments and completely different prognosis. The purpose of this study is to develop an Artificial Intelligence (AI) which can determine imaging features in a normal x-ray to distinguish osteomyelitis from Ewing sarcoma. Is it possible to develop an AI that can support a radiological assessment of orthopaedic surgeons, so that unnecessary time between presentation and referral to a specialized cancer center can be saved? Can we allocate with only an x-ray, the cheapest and most accessible of all diagnostic imaging instruments, a young cancer patient?

Methods: We collected 113 radiographs from our sarcoma center at Klinikum rechts der Isar (47 ES, 66 acute and chronic osteomyelitis) from patients younger than 18 years with no specific localization. We developed an algorithm around these x-ray images, which could distinguish radiological features of these entities (Figure 1 [Fig. 1]).

In order to grasp low level features, a deep version of a classification architecture was implemented (ResNet152). To tackle the limited and unbalanced amount of data, extensive data augmentation as well as a method to weight the losses accordingly supported the algorithm.

We applied a data split of 70%, 20%, 10% for training, validation and testing, respectively. To provide statistical significance and robustness, the data from a single patient was either part of training, validation or testing. An additional 5-fold cross validation supported this task, while random chosen test data for final evaluation remained untouched.

Results and Conclusion: The algorithm achieved an accuracy of 81% on validation and 77% on test data in differentiation of ES from osteomyelitis. Due to the utilization of a pretrained network, the results showed considerable robustness. Still, an even higher accuracy is to be expected with more data.

We think that this AI algorithm can be a valuable support for any physician involved in the differential diagnosis process of musculoskeletal Tumors. This allowing a minimal loss of time between diagnosis and specific treatment.

As a next step we will collect more data to improve the accuracy and to provide a valuable high precision tool.