DIAGNOSING LIVER'S LESIONS FOR MEDICAL ANALYSIS

Abstract

Machine learning (ML) and computer vision techniques have grown rapidly due to their automation, suitability, and ability to generate astounding results, especially the convolutional neural network (ConvNet). In this thesis, we survey the critical studies published between 2014 and 2020, showcasing the different ML algorithms researchers have used to segment the liver, hepatic-tumors, and hepatic-vasculature structures. Following that, and stemming from the surveyed literature, we propose our methodology that tackles a famous dataset named Medical Segmentation Decathlon Challenge Task 8: Hepatic Vessels (MSDC-T8), which has all the liver tissues manually segmented (liver, tumors, and vessels). This dataset is also considered the largest publicly available dataset for tackling the liver tissues delineation challenge. It encapsulates a total of 443 contrast-enhanced computerized tomography (CE-CT) scans, where the ground-truth liver masks are available for all the volumes, and the tumors and vessels segmentations are known for 303 of them. Correspondingly, this methodology is applied for each tissue of interest (TOI), as we first tackle the liver segmentation, followed by the tumors and vessels segmentation in parallel. We compare different training environment parameters via famously used percentile and distance metrics. In our results, our liver segmentation ConvNet has surpassed the state-of-the-art performance by scoring a Dice of 98.12% on the MSDC-T8. Moreover, the tumors and vessels segmentation ConvNets compete with the state-of-the-art, scoring a ~60% Dice for tumors' segmentation task (with the best model scoring 65.95%) and ~50% for the vessels' segmentation task (with the best model scoring 51.94%). Finally, when all the masks are segmented, a 3D interpolation is created for the liver (showing its tumors and blood vessels) and is exported into both .obj and .mtl files, which are 3D printing friendly. To manifest the usefulness of our work, we create a user-friendly desktop application that allows clinicians to import CT scans of selected patients. This desktop application's output is the aforementioned 3D interpolated object represented by the .obj and .mtl files.