Automated Strategies in Multimodal and Multidimensional Ultrasound Image-based Diagnosis

Medical ultrasonography is an effective technique in traditional anatomical and functional diagnosis. However, it requires the visual examination by experienced clinicians, which is a laborious, time consuming and highly subjective procedure. Computer-aided diagnosis (CADx) have been extensively used in clinical practice to support the interpretation of images; nevertheless, current ultrasound CADx still entails a substantial user-dependency and are unable to extract image data for prediction modelling. The aim of this thesis is to propose a set of fully automated strategies to overcome the limitations of ultrasound CADx. These strategies are addressed to multiple modalities (B-Mode, Contrast-Enhanced Ultrasound-CEUS, Power Doppler-PDUS and Acoustic Angiography-AA) and dimensions (2-D and 3-D imaging). The enabling techniques presented in this work are designed, developed and quantitively validated to efficiently improve the overall patients’ diagnosis. This work is subdivided in 2 macro-sections: in the first part, two fully automated algorithms for the reliable quantification of 2-D B-Mode ultrasound skeletal muscle architecture and morphology are proposed. In the second part, two fully automated algorithms for the objective assessment and characterization of tumors’ vasculature in 3-D CEUS and PDUS thyroid tumors and preclinical AA cancer growth are presented. In the first part, the MUSA (Muscle UltraSound Analysis) algorithm is designed to measure the muscle thickness, the fascicles length and the pennation angle; the TRAMA (TRAnsversal Muscle Analysis) algorithm is proposed to extract and analyze the Visible Cross-Sectional Area (VCSA). MUSA and TRAMA algorithms have been validated on two datasets of 200 images; automatic measurements have been compared with expert operators’ manual measurements. A preliminary statistical analysis was performed to prove the ability of texture analysis on automatic VCSA in the distinction between healthy and pathological muscles. In the second part, quantitative assessment on tumor vasculature is proposed in two automated algorithms for the objective characterization of 3-D CEUS/Power Doppler thyroid nodules and the evolution study of fibrosarcoma invasion in preclinical 3-D AA imaging. Vasculature analysis relies on the quantification of architecture and vessels tortuosity. Vascular features obtained from CEUS and PDUS images of 20 thyroid nodules (10 benign, 10 malignant) have been used in a multivariate statistical analysis supported by histopathological results. Vasculature parametric maps of implanted fibrosarcoma are extracted from 8 rats investigated with 3-D AA along four time points (TPs), in control and tumors areas; results have been compared with manual previous findings in a longitudinal tumor growth study. Performance of MUSA and TRAMA algorithms results in 100% segmentation success rate. Absolute difference between manual and automatic measurements is below 2% for the muscle thickness and 4% for the VCSA (values between 5-10% are acceptable in clinical practice), suggesting that automatic and manual measurements can be used interchangeably. The texture features extraction on the automatic VCSAs reveals that texture descriptors can distinguish healthy from pathological muscles with a 100% success rate for all the four muscles. Vascular features extracted of 20 thyroid nodules in 3-D CEUS and PDUS volumes can be used to distinguish benign from malignant tumors with 100% success rate for both ultrasound techniques. Malignant tumors present higher values of architecture and tortuosity descriptors; 3-D CEUS and PDUS imaging present the same accuracy in the differentiation between benign and malignant nodules. Vascular parametric maps extracted from the 8 rats along the 4 TPs in 3-D AA imaging show that parameters extracted from the control area are statistically different compared to the ones within the tumor volume. Tumor angiogenetic vessels present a smaller diameter and higher tortuosity. Tumor evolution is characterized by the significant vascular trees growth and a constant value of vessel diameter along the four TPs, confirming the previous findings. In conclusion, the proposed automated strategies are highly performant in segmentation, features extraction, muscle disease detection and tumor vascular characterization. These techniques can be extended in the investigation of other organs, diseases and embedded in ultrasound CADx, providing a user-independent reliable diagnosis.