A Deep Learning Approach for Segmentation of Ovarian Adnexal Masses

Ovarian cancer is the eighth most widespread cancer in the world among women. Due to the absence of a specific symptomatology and the lack of a defined screening protocol, this disease is usually diagnosed at an advanced stage, leading to the increase in the corresponding mortality rate. At present, transabdominal sonography (TAS) and transvaginal sonography (TVS) are generally recognised as the main diagnostics techniques for the first identification of the neoplasm. However, it is well known that their diagnostic effectiveness can be seriously compromised by the intrinsic noisy and operator-dependent nature of ultrasound images. In addition, the identification of ovarian structures is a time-consuming task in clinical practice, almost always repetitive and prone to errors when manually performed by medical doctors. New diagnostic approaches based on artificial intelligence algorithms have started to be investigated for the automatic detection, segmentation and classification of medical images, aiming at the development of Computer Aided Diagnosis (CAD) models. Although these protocols are commonly employed in the examination of other image acquisition modes, such as CT or MRI, their application is also expanding to ultrasound images. Nowadays, the majority of artificial intelligence algorithms in gynaecological ultrasound is mostly focused on the classification of ovarian mass types. Indeed, despite many advances have already been made to identify anatomical structures of the same district, such as ovarian follicles, very few has been done concerning ovarian mass segmentation. An increasingly widespread idea in medical imaging is that the automatic segmentation of ovarian masses could be consistently helpful for the development of CAD systems. For instance, the emerging field of radiomics could strongly benefit from the development of such segmentation protocols: the identification of ROIs within medical images is a mandatory step for the extraction of quantitative descriptors useful for tumor discrimination. The purpose of this thesis project is the development of an automatic algorithm that deals with the segmentation of ovarian masses. Given the success of Deep Learning models in the medical field, and, in particular, of fully convolutional neural network (FCNN) in image segmentation tasks, the learning model proposed in this work makes use of a modified U-net network with MobileNetV2 as encoding block and simple transpose-deconvolutional-based upsampling as decoding block. Being one of the few attempts to the segmentation of ovarian masses, its future potential has been here evaluated on an easier segmentation task, that is the identification of cysts of unilocular serous type. The model showed to be perfectly suitable to perform the proposed task and confirmed the improvement over the current state-of-the-art. Its performance have been further improved with both the introduction of Data Augmentation and a refining post-processing stage. The success of this segmentation algorithm encourages as a next step its subsequent application to the segmentation of histotypes with more complex morphology than those covered in the present study. Another possible implementation could be aimed at multiclass segmentation of different cystic components. The automatic identification of these substructures would contribute to the development of a more interpretable and less ``black-box'' algorithm for differential diagnostics of ovarian lesions.