St Andrew's War Memorial Hospital ยท Spring Hill, Brisbane For referrers 07 3063 5500
Menu
ShouldersRehabilitationShoulder AnatomyShoulder Injuries and ConditionsShoulder SurgeryVideosSurgeries and RehabilitationACJ ExcisionAcromioplastyAll Surgeries and RehabilitationAnatomic Total Shoulder ReplacementArthroscopic Bankart RepairBiceps TenodesisBrachial Plexus ExplorationDistal Biceps RepairExcision of CalciumFrozen Shoulder ReleaseLatarjet and Labral RepairORIF Scapula, Clavicle and HumerusOs Acromiale SurgeryPec Major RepairRemplissageReverse Total Shoulder ReplacementRotator Cuff RepairScapula ResectionStabilisationSternoclavicular Joint StabilisationInformationAccommodationConsultationConsultation FAQsSurgery FAQsAboutDr Kenneth CutbushInside the Shoulder PodcastPhysiotherapyPublicationsQualificationsResearchShoulder FellowshipStaffContactNews

Publications ·28 May 2024

New deep learning framework for detecting abnormalities in shoulder X-ray images

Deep learning framework for detecting abnormalities in shoulder X-ray images
Deep learning framework for detecting abnormalities in shoulder X-ray images

The study

This research, published in PLOS ONE, concerns the development of a trustworthy deep learning framework for detecting abnormalities in X-ray images of the shoulder.

Deep learning framework for detecting abnormalities in shoulder X-ray images

What deep learning on an X-ray means

Deep learning is a method in which a computer model is shown a large number of labelled examples and learns, by adjusting itself repeatedly, which patterns in an image go with which label. Applied to shoulder radiographs, the task is to flag images that look abnormal so that they can be looked at carefully. The model is not making a diagnosis and does not replace the clinician reading the film; the proposition is that it can direct attention. PLOS ONE is a peer-reviewed journal, so the work was assessed by independent researchers in the field before publication.

The problem of being confidently wrong

The focus of the work is building AI models that are more reliable in clinical use. The specific problem it addresses is that of models being "confidently incorrect": producing a wrong answer while reporting high certainty about it.

This matters more in medicine than in most applications. A model that says "I am not sure" can be referred on to a human. A model that is wrong and sure of itself is harder to catch, and in a setting where the output may influence a clinical decision, that is the failure mode that causes harm. Addressing it is a prerequisite for using such tools responsibly rather than an optional refinement.

The practical consequence is that a useful model has to report uncertainty that corresponds to how often it is actually right. A system that reserves its high confidence for the cases it reliably gets correct, and declares doubt on the rest, can be built into a workflow, because the cases it is unsure about can be routed to a person. That is what the word trustworthy is doing in the title of the paper.

First author

The first author is Dr Laith Alzubaidi, a postdoctoral researcher at Queensland University of Technology, who holds a master's degree from the University of Missouri and has published more than 40 refereed research papers.

Dr Laith Alzubaidi, first author

Related information