An innovative annotation tool for selecting Regions of Interest (ROI) in contrast-enhanced imaging ultrasonography (CEUS) of Hepatocellular carcinoma (HCC)

Introduction

CEUS is a safe and cost-effective imaging technique which allows a real-time evaluation of focal liver lesions (FLL). It has become a fundamental tool in HCC surveillance and diagnosis. Nevertheless, some concerns have risen in the past on its diagnostic accuracy, in particular in its ability to discriminate between HCC and small (<3 cm) Intraepatic cholangio-Carcinoma (ICC). Recent advances in AI-driven tools for medical imaging have demonstrated the potential of greatly improving accuracy. In ultrasonography, they have been proposed to perform diagnosis from CEUS images alone or in combination with Bi-modal ones: the former can provide information about the contrast pattern; the latter enhances structural characteristics.Two challenges arise in the use of machine learning in the analysis of liver ultrasound images. First, the lack of large public datasets, which makes it difficult to train deep learning models, capable of extracting powerful features, optimized for the task: Radiomics, represents an effective technique for extracting powerful quantitative features without the need of a training dataset. Second, the need for complete and standardized annotations, crucial for effective model training.

Aims

To build an annotation software capable of selecting ROIs in CEUS imaging and of extracting Radiomics features from both CEUS and B-Mode views, with the intention to build an automated detection software protocol for FLL diagnosis.

Methods

A stand-alone tool for annotating CEUS exams and extracting Radiomics features from the annotated ROIs is proposed. It inputs a CEUS exam stored in a WMV or a DICOM file, which includes both B-mode and CEUS views in split screen manner. It allows the radiologist to select the more suitable frame to be annotated within the exam. Two automatisms are introduced in the annotation procedure: first, the tool is capable of replicating in real-time the annotation performed by the annotator on one view, onto the other one. In addition, it is capable of automatically interpolating contour points of the lesion, in case the annotator will not perform a complete annotation. Once annotation is performed, both annotated CEUS and B-Mode views are prepared for labeling and feature extraction. Labeling concerns both the selection of the phase the frame belongs to, i.e., arterial, portal or late, and the type of annotated lesion. Feature extraction instead is performed through Radiomics: to this aim, the annotated views are automatically saved into two distinct NifTI files, from which histogram-based, shape-based and texture based Radiomics features are automatically extracted through PyRadiomics and saved into a specific folder for that exam. A total number of 102 features are automatically extracted and saved from each view.

Results

The proposed tool has been positively evaluated by radiologists from the University of Salerno, because of its ease of use and its ability to automatically reproduce on one view, the annotations manually made on the other one.

Conclusions

In this work, a stand-alone tool for annotating multi-view liver ultrasound images is proposed. The next step is to use it to create a complete labeled dataset exploitable by AI methods from single or multi view liver lesion analysis.