Expert-Level Reliability of Automated Skin Ultrasonography Segmentation

The treatment of skin diseases such as skin cancer and chronic inflammatory skin disorders are highly dependent on the clinical assessment of the shape and size of lesions, for which high frequency ultrasound (HFUS) has shown promise. However, the interpretation of ultrasound images can be challenging even for experts. Therefore, automatic segmentation of different skin structures and lesions on HFUS images could prove to be useful. The current work uses a deep neural network (DNN) to automatically segment different skin structures and mark lesions on HFUS images, and evaluates its performance compared to human physician experts. A portable HFUS imaging device with optical image guidance was used to acquire a dataset of 1691 skin ultrasound samples. Two labellers annotated the dataset, and it was split into training and testing subsets. The self-configuring nnU-Net DNN framework was used for segmentation. Performance was evaluated by comparing the DNN segmentation of all skin structures with the labels of a test set, and by comparing the annotations of an attending radiologist and dermatologist with respect to the model and each other using the Dice score and estimation error (standard deviation) of the maximal lesion depth (hereon "lesion depth error"). The developed DNN accurately segmented the epidermis, dermis, subcutis and lesion, with Dice coefficients of 0.863, 0.922, 0.962, and 0.728, respectively. The agreement between the two human experts in terms of the Dice coefficient and lesion depth error was 0.726, 0.42 mm, respectively. The model provided comparable accuracy, with Dice coefficients of 0.731 (radiologist) and 0.733 (dermatologist), and lesion depth errors of 0.38 mm (radiologist) and 0.19 mm (dermatologist). The results validate the performance of the automated skin segmentation model and highlight the need to compare performance with inter-expert reliability. The use of the machine learning model as an interpretation tool for medical experts may in the future help determine the shape and size of skin lesions, which has a direct impact on their management.