The role of doppler artifacts in color score assessment. Algorithms to remove artifacts from doppler signal and to evaluate vascularization of adnexal lesions in diagnostic ultrasound.

Ovarian cancer accounts for 3.3% of all cancers in women worldwide but it is responsible for approximately half of the deaths related to gynecological cancer. The low survival rate is primarily due to the difficulty in diagnosing this neoplasm at early stages, and this occurs because early ovarian cancer is usually asymptomatic. Moreover, an adequate discrimination between benign and malignant lesions is a key point for a correct management of the patient. Ultrasound is the most used method for assessment of adnexal lesions due to its non-invasiveness, low cost and widely diffusion. Specifically, Color Doppler and Power Doppler imaging techniques are used to differentiate between benign and malignant tumors by evaluating the vascularization within the adnexal mass. According to the International Ovarian Tumor Analysis group, vascularization can be described by a scoring system, i.e. the Color Score (CS), assigned by clinicians when inspecting a Color or Power doppler examination. CS is 1 when no blood flow is present within the lesion, 2 with minimal vascularization, 3 with moderate blood flow, 4 if there is a high vascularization. However, despite CS being a good predictor of malignancy, the estimation of the color content within a lesion is based on the subjective evaluation of clinicians. Moreover, Doppler techniques are characterized by the presence of several artifacts that can influence the assignment of the CS making more difficult for clinicians to interpret the flow information. Consequently, two algorithms were developed with the aim of removing artifacts and easing clinicians’ evaluation: ·??Pixel-based denoising algorithm: it removes doppler artifacts based on the temporal persistence of the colored (doppler) pixels within the video, assuming that artifacts are not persistent. However, it only relies on the exact pixel correspondence in a frame sequence, with some real activations flagged as artifacts because they shift through frames due to the probe moving during the acquisitions. ·??Connected components-based denoising with component-tracking: it overcomes the limit of the pixel-based denoising by considering both temporal and spatial persistence: it relies on connected components, rather than single pixels. It suppresses artifacts while tracking the activations’ clusters – connected components of colored pixels – during the video. The algorithms were developed on 101 ovarian cancer cases from two hospitals. To assess the effect of the artifact-removal algorithms, I trained a decision tree to predict CS based on the doppler estimation obtained on both non processed and denoised doppler videos of 106 ovarian cancer cases. The doppler was estimated as the number of colored pixels present within the lesion, whereas CSs were assigned by a panel of 6 expert clinicians. Three experiments were conducted: in the first, the algorithm was trained and tested on non processed videos; in the second applying the pixel-based denoising; in the third applying the component-tracking denoising. The results show that applying the artifact-removal algorithms generally improved classification performances with respect to non processed videos. Moreover, performance is higher with the pixel-based algorithm compared to the tracking one. I conclude that removing artifacts provides a more accurate estimate of doppler signal for color score assessment. Further studies are needed to improve the tracking approach, for example by classifying the artifact from the signal connected components.