Classifying washout in contrast-enhanced ultrasound examinations of focal liver lesions using deep learning

Hintergrund Contrast-enhanced ultrasound (CEUS) is a reliable tool to diagnose liver lesions which appear ambiguous in B-mode ultrasound. However, interpretation of the dynamic contrast sequences can be challenging. This work investigates a deep learning-based image classifier to determine the feature washout from CEUS acquisitions. Washout assessment is important to distinguish benign and malign lesions, therefore its automation could serve as starting point for a decision support system.

Methoden 500 CEUS acquisitions of liver lesions were collected at two clinical sites. Data was annotated by one radiologist per site by contouring the lesion and a parenchyma area in one representative frame of each contrast phase. Additionally, a report regarding lesion appearance in the different phases as well as diagnostic decision was filled out. After excluding data with incomplete annotations, 481 liver lesions were available.

Data was converted into a 2D representation to serve as input to the ConvNeXt architecture which classified washout into categories [No Washout, Washout, Neither]. Experiments include strategies to select additionally frames beside the annotated ones: a correlation-threshold based frame selection as well as selection of frames from similar breathing cycles.

Ergebnisse The radiologists' specifications were compared with the results of the algorithms in a confusion matrix ([Fig. 1]).

Fig. 1  Confusion matrices for the three investigated frame selection strategies on the cross-validation data. Data is normalized per row and given in percent. The networks performing best for breathing and correlation threshold frame selection are shown.

Using only the three annotated frames, a balanced accuracy of 81.0% is reached. While the breathing cycle-based approach achieves 81.4%, the threshold-based approach performed best (84.0%). For the latter two, an optimisation regarding their correlation threshold was performed.

Schlussfolgerung Automatic washout classification from CEUS sequences is feasible in principle, providing the opportunity to support in the distinction of malignant and benign liver lesions. Improvements in performance were observed when integrating more frames, highlighting the importance to take the time-dependent characteristics of CEUS into account.

 

Scroll to top

Publikationsverlauf

Artikel online veröffentlicht:
16. Oktober 2025