Using Machine Learning to Diagnose Misaligned CT Scans

Sheridan C. Perry, Embry-Riddle Aeronautical University
Matthew Folkman, Rainbow Babies and Children’s Hospital
Takara O'Brien, Embry-Riddle Aeronautical University
Lauren A. Wilson, Embry-Riddle Aeronautical University
Eric Coyle
Raymond W. Liu, Rainbow Babies and Children’s Hospital
Charles T. Price, International Hip Dysplasia Institute
Victor Huayamave, Embry-Riddle Aeronautical University

Abstract

The usage of machine learning has grown exponentially in recent years; However, its applicable uses for medical diagnosis are still in an early stage. Conditions such as Developmental Dysplasia of the Hip (DDH), Cerebral Palsy (CP), and Femoracetabular Impingement (FAI) rely heavily on imaging techniques such as Ultrasound and Computed Tomography (CT) scans. Radiologists use multiple manually computed metrics using these images to diagnose conditions. This is time-intensive and requires an aligned image to get accurate diagnoses. The proposed application uses a deep learning detection algorithm to assist in the metric computation process. The algorithm is implemented using MATLAB R2023A and is trained on CT data gathered from 60 healthy participants. The algorithm performed well on images aligned according to the standard anteroposterior alignment used for radiological measurement. However, the variance of the metrics computation significantly increases when faced with severe misalignment in the craniocaudal or mediolateral axes. Additional algorithm improvements must be made to overcome this increased variance.

 

Using Machine Learning to Diagnose Misaligned CT Scans

The usage of machine learning has grown exponentially in recent years; However, its applicable uses for medical diagnosis are still in an early stage. Conditions such as Developmental Dysplasia of the Hip (DDH), Cerebral Palsy (CP), and Femoracetabular Impingement (FAI) rely heavily on imaging techniques such as Ultrasound and Computed Tomography (CT) scans. Radiologists use multiple manually computed metrics using these images to diagnose conditions. This is time-intensive and requires an aligned image to get accurate diagnoses. The proposed application uses a deep learning detection algorithm to assist in the metric computation process. The algorithm is implemented using MATLAB R2023A and is trained on CT data gathered from 60 healthy participants. The algorithm performed well on images aligned according to the standard anteroposterior alignment used for radiological measurement. However, the variance of the metrics computation significantly increases when faced with severe misalignment in the craniocaudal or mediolateral axes. Additional algorithm improvements must be made to overcome this increased variance.