Segmentation of skin layers in ultrasound images using a crowdsourcing and deep learning-based system

Ultrasonography (US) has demonstrated many advantages in the detection, characterization, and monitoring of different diseases. Through high frequency probes, it is possible to visualize and characterize the anatomical layers, such as tissues, which can constitute a very helpful supporting tool in several procedures, e.g., surgeries. However, the visual identification of tissues in this type of image is still a challenge for some professionals. In this work, we evaluate deep learning (DL) segmentation algorithms for the tissue segmentation task in US. Moreover, we briefly assess whether their performance can be improved by including a crowdsourcing step. In order to perform the segmentation task, different segmentation models are trained, as posteriorly, the crowdsourcing step is included. The proposed approach is composed of the following steps: 1 - automatic segmentation using a deep learning algorithm; 2 - crowd evaluation and correction of the results. In order to perform step 1, different segmentation models are trained. The second step includes a visual interface where users can: a) validate the quality of the automatic segmentation; or b) correct the segmentation whether the DL result is inconsistent. All users are scored, denoting the quality of their annotations, considering the manual annotations provided by an expert for a small group of images. In order to evaluate the method and compare it to a DL algorithms alone, a total of 100 US images were used. Our experiments show that the inclusion of crowdsourcing significantly improved the performance of the tissue segmentation task compared to using the DL models alone. The performance of our method demonstrated the feasibility of applying this type of solution for the considered problem of segmenting tissues in US facial images. Moreover, the results suggest that this tool can be employed as an auxiliary tool in oral procedures.