Deep Learning Has Human Touch in Diagnosing an Anterior Cruciate Ligament Tear

By Nick Klenske

While the use of deep learning (DL) for detecting disease on medical images is a hot area in radiology, one researcher is using the technology in a new area: to create a fully-automated prediction model for detecting anterior cruciate ligament (ACL) tears on MR images.

Kijowski

In a Tuesday session, researcher Richard Kijowski, MD, from the Department of Radiology at the University of Wisconsin-Madison, discussed his study evaluating the ability of DL — a type of machine learning — to detect musculoskeletal (MSK) disorders on MRI.

Dr. Kijowski and colleagues trained the DL machine to determine the presence or absence of an ACL injury in 125 patients with surgically-confirmed ACL tears (100 in training group, 25 in validation group) and 125 patients without ACL tears (100 in training group, 25 in validation group).

The DL method was then evaluated on a hold-out test group of an additional 100 subjects, half with and half without surgically-confirmed ACL tears. For comparison purposes, an experienced MSK radiologist, an MSK radiology fellow and three radiology residents with varying levels of clinical experience also determined the presence or absence of an ACL tear in the same 100 subjects evaluated by the DL machine.

DL Results Similar to Human Readers

Results showed that the DL method had similar sensitivity and specificity as the human readers for determining the presence and absence of an ACL tear (96 percent sensitivity, 96 percent specificity).

"These results show that deep learning methods offer a diagnostic performance similar to experienced radiologists," Dr. Kijowski said.

According to Dr. Kijowski, DL is a powerful tool with great potential for helping radiologists interpret medical images. For example, as the detection time of the machine is highly efficient in the order of seconds, it can be quite useful as a rapid screening method for providing an immediate interpretation of abnormal results. Furthermore, the machine is not influenced by errors due to inexperience, distraction, and fatigue associated with human interpretation. Nevertheless, Dr. Kijowski noted there are limitations to the current DL method for detecting ACL tears on MRI. "Although the DL machine made a few excellent diagnoses, there were cases of false positive and negative interpretations that were readily detectable by human readers," he said.

For this reason, additional work is needed before the ACL tear detection system can be implemented in clinical practice. Specifically, much larger annotated image datasets are required to train the system, and larger hold-out test datasets containing images from different types of scanners are needed for validation.