Ultrasound Radiomics Correlating With Clinical Markers for Enhanced Detection of Placenta Accreta Spectrum.

OBJECTIVES

(i) To develop an accurate and robust ultrasound-based quantitative index for placenta accreta spectrum (PAS) diagnosis by amalgamating texture features from B-mode and color Doppler imaging. (ii) To test the correlation of this index with established sonographic markers for PAS.

METHODS

In this retrospective study, we collected 2106 texture features extracted from the midsagittal placental view of mid-trimester B-mode and Doppler ultrasound images. These images were acquired during ultrasound examinations conducted between 24 and 34 wk gestation from patients diagnosed with placenta previa and at least one prior uterine surgery risk factor for PAS at a tertiary center in Toronto, Canada. Three distinct models were developed: one using B-mode data (n = 174), another using color Doppler data (n = 98), and a third integrating both modalities (n = 98). Integrated features, derived from weighted z-scores, were employed to generate a quantitative index for detecting PAS. A feature selection pipeline was implemented, combining linear discriminant analysis, extra trees regression and recursive feature elimination. The pipeline identified 15-feature subsets for both B-mode and Doppler models, and a 20-feature subset for the multimodal model, both of which were used for training. The predictive performance of each model (presence/absence of PAS) was assessed using five-fold cross-validation and tested on a separate hold-out test set. Subsequently, the optimized quantitative metrics were examined for any statistically significant correlations with established sonographic clinical markers.

RESULTS

The five-fold cross-validation accuracies across the developed B-mode, color Doppler and multimodal models were 88.7% (±5.3), 85.1% (±9.1) and 90% (±6.7), respectively. All eight of the clinical disease markers evaluated from each ultrasound image were determined to be discriminative of the quantitative index generated from B-mode and multimodal models, and six of the eight markers were discriminative of the index generated from Doppler alone.

CONCLUSION

Combined B-mode and Doppler ultrasound-based radiomics can accurately detect PAS in patients with placenta previa and provide a surgical risk factor for PAS using simple and computationally inexpensive operations. The rank correlations between B-mode and color Doppler quantitative indices and their corresponding clinical markers highlight that this multimodal approach captures crucial diagnostic information that may be missed by individual models. All models exhibit notably higher accuracy than projected clinical standards and suggest their potential as aids for clinical decision-making at routine mid-trimester scans.