Classification and risk assessment of ovarian-adnexal lesions using parametric and radiomic analysis of co-registered ultrasound-photoacoustic tomographic images

Abstract

Ovarian-adnexal lesions are conventionally assessed with ultrasound (US) under the guidance of the Ovarian-Adnexal Reporting and Data System (O-RADS). However, the low specificity of O-RADS results in many unnecessary surgeries. Here, we use co-registered US and photoacoustic tomography (PAT) to improve the diagnostic accuracy of O-RADS. Physics-based parametric algorithms for US and PAT were developed to estimate the acoustic and photoacoustic properties of 93 ovarian lesions. Additionally, statistics-based radiomic algorithms were applied to quantify differences in the lesion texture on US-PAT images. A machine learning model (US-PAT KNN model) was developed based on an optimized subset of eight US and PAT imaging features to classify a lesion as either cancer, one of four subtypes of benign lesions, or a normal ovary. The model achieved an area under the receiver operating characteristic curve (AUC) of 0.969 and a balanced six-class classification accuracy of 86.0 %.

Keywords

• Ovarian-adnexal lesions
• O-RADS
• Co-registered ultrasound-photoacoustic tomography
• Multiparametric ultrasound
• Radiomics