AbstractBackgroundFrank’s sign (FS) is a diagonal crease in the ear lobe extending from the tragus across the lobule to the rear edge of the auricle. Many of previous research have studied that FS is associated with the risk of cardiovascular, cerebrovascular diseases, cognitive impairment and aging. However, existing studies have lack of work in the application of FS. Meanwhile, cutting edge deep learning segmentation techniques allow to build large‐scale trainable models that have the capacity to learn the optimal features with great speed and accuracy. This study is the first to analyze 3D Nifti FS dataset by applying a deep learning approach and developing the automatic detection algorithm of FS (ADAFS).MethodWe used 700 brain MRI dataset and manually segmented FS masks of 550 older adults with FS. 400 out of 550 datasets were used into the developing ADAFS. Rest of them and 150 older adults without FS served to validate the ADAFS. The preprocessed 3D datasets were fed into the candidate models including 3D U‐net, 3D Attention U‐net, 3D Nested U‐net, and 3D USE‐net in order to train ADAFS (figure 1). We adapted 5‐fold cross‐validation test which can estimate the general effectiveness of an algorithm on an independent dataset. Furthermore, we classified the FS‐detected mask in 4 cases; left lateral, right lateral, bi‐lateral and no FS.ResultOn development of ADAFS, Nested U‐net out of 4 different models showed the highest average dice similar coefficient (DSC; 0.714 ± 0.160) across test folds. Given validating the developed ADAFS, the multi‐label classification and segmentation results of receiver operating characteristic curve (ROC) analysis were below: accuracy = 0.907, average area under ROC curve = 0.940, and DSC = 0.736 ± 0.172.ConclusionADAFS may support the faster decisions of FS presence, and considerably similar FS segmentation performance. This may serve a variety of FS shape descriptors and further study including an association of cognitive aspects and them.
Read full abstract