Pattern recognition methods for detection of ovarian cysts papillary projections in sonographic videos

Ovarian cancer is one of the leading causes of tumor related deaths in women and it has the highest mortality rate among gynecological cancer diseases. The reason why this type of cancer is among the deadliest ones is due to the fact that the absence of specific symptoms leads to a delay in diagnosis. Moreover, ovarian cancer diagnosis is often performed with sonography (ultrasound, US), which is considered the gold standard examination, but shows high user-dependance. Finding a way to get an early diagnosis is fundamental to drastically reduce mortality. According to the international guidelines adopted by the medical community, the presence of papillary projections within ovarian cysts is a discriminant feature between benign and malignant conditions. However, the detection of the papillary projections in US videos is a complex task and affected by variability in observer interpretation. Indeed, papillary projections share with other structures, such as ultrasound artifacts, similar sonographic features. These structures could be a source of misclassification. The aim of this study was to find a predictive model able to support sonographic evaluation and detect papillary projections within US videos. For this purpose, a set of US videos, completed with pathology reports, was used to extract US images that were subdivided into training and testing data for a classification algorithm. Semi- automatic segmentation step was performed to extract regions of interest from training US images. Image pre- processing and data augmentation processes were performed in order to increase the heterogeneity associated with the small number of different clinical cases used in the study and to process input data from different ultrasound devices. Statistical and textural features were extracted to characterize the echostructures of papillary projections and ultrasound artifacts, in order to build a supervised machine learning model based on the Linear SVM algorithm. The model was able to classify within US images papillary projections and ultrasound artifacts with an accuracy of 97% on training data, 94% on validation data and 72% on test data. With validation data, sensitivity and specificity were above 91% and 94%, respectively. A further analysis of the effective ability of the model to detect papillary projections within US videos and related clinical implications were provided.