Interpretable Machine Learning for Characterization of Focal Liver Lesions by Contrast-Enhanced Ultrasound

Abstract

This work proposes an interpretable radiomics approach to differentiate between malignant and benign focal liver lesions (FLLs) on contrast-enhanced ultrasound (CEUS). Although CEUS has shown promise for differential FLLs diagnosis, current clinical assessment is performed only by qualitative analysis of the contrast enhancement patterns. Quantitative analysis is often hampered by the unavoidable presence of motion artifacts and by the complex, spatiotemporal nature of liver contrast enhancement, consisting of multiple, overlapping vascular phases. To fully exploit the wealth of information in CEUS, while coping with these challenges, here we propose combining features extracted by the temporal and spatiotemporal analysis in the arterial phase enhancement with spatial features extracted by texture analysis at different time points. Using the extracted features as input, several machine learning classifiers are optimized to achieve semiautomatic FLLs characterization, for which there is no need for motion compensation and the only manual input required is the location of a suspicious lesion. Clinical validation on 87 FLLs from 72 patients at risk for hepatocellular carcinoma (HCC) showed promising performance, achieving a balanced accuracy of 0.84 in the distinction between benign and malignant lesions. Analysis of feature relevance demonstrates that a combination of spatiotemporal and texture features is needed to achieve the best performance. Interpretation of the most relevant features suggests that aspects related to microvascular perfusion and the microvascular architecture, together with the spatial enhancement characteristics at wash-in and peak enhancement, are important to aid the accurate characterization of FLLs.

Fig. 1.

Flowchart describing the processing and machine learning pipelines. (a) Side-by-side view of B-mode and CEUS. (b) Manual segmentation of the lesion on the B-mode image. (c) Automatic definition of the ROI based on the location of the manually drawn lesion. (d) Extraction of spatiotemporal features (using all frames) and texture features at wash-in, peak, and wash-out frames. (e) Extraction of summary statistics from ROI and feature filtering for dimensionality reduction; N represents the number of selected features at each step, while M represents the number of samples. (f) Repeated nested k-fold cross-validation procedure for hyperparameter tuning (inner loop, yellow) and performance evaluation (outer loop, orange).

Fig. 2.

Two examples of average TIC (TIC_mean) obtained from the lesion ROI (blue stars), together with the straight-line fit in the wash-in and wash-out (orange solid lines). The times at which the wash-in, peak, and wash-out frames were selected are indicated by dashed vertical lines.

Fig. 3.

(a)–(e) Examples of parametric maps obtained for one benign and (f)–(l) one malignant lesion: (a) and (f) side-by-side view of B-mode and CEUS at the reference frame, with manually delineated lesion and the analysis ROI highlighted in blue and red, respectively; (b) and (g) spatiotemporal feature “Coherence,” (c) and (h) spatiotemporal feature “peak time”; (d) and (i) texture feature “Global Kurtosis” at wash-in; (e) and (l) texture feature “GCLM Energy” at peak (visualized in logarithm scale).

Fig. 4.

Analysis of the number of features required to optimize the bACC for each model. The optimal number for each classifier is highlighted with a black circle.

Fig. 5.

Normalized PFI of the top 10 features for the LR, SVM, and sVC classifiers. The percentage of times that a feature was chosen is given by the text on top of each bar and additionally indicated by the shading of the bars. For the texture features, the prefixes “WiIm,” “PkIm,” and “WoIm” indicate features extracted at wash-in, peak intensity, and wash-out, respectively. For all features, the suffixes “median,” “iqr,” and “skew” indicate the median, interquartile range, and skewness, respectively, extracted over the ROI.