Comparison of Intraoperative Ultrasound B-Mode and Strain Elastography for the Differentiation of Glioblastomas From Solitary Brain Metastases. An Automated Deep Learning Approach for Image Analysis

Abstract

Background: The differential diagnosis of glioblastomas (GBM) from solitary brain metastases (SBM) is essential because the surgical strategy varies according to the histopathological diagnosis. Intraoperative ultrasound elastography (IOUS-E) is a relatively novel technique implemented in the surgical management of brain tumors that provides additional information about the elasticity of tissues. This study compares the discriminative capacity of intraoperative ultrasound B-mode and strain elastography to differentiate GBM from SBM.

Methods: We performed a retrospective analysis of patients who underwent craniotomy between March 2018 to June 2020 with glioblastoma (GBM) and solitary brain metastases (SBM) diagnoses. Cases with an intraoperative ultrasound study were included. Images were acquired before dural opening, first in B-mode, and then using the strain elastography module. After image pre-processing, an analysis based on deep learning was conducted using the open-source software Orange. We have trained an existing neural network to classify tumors into GBM and SBM via the transfer learning method using Inception V3. Then, logistic regression (LR) with LASSO (least absolute shrinkage and selection operator) regularization, support vector machine (SVM), random forest (RF), neural network (NN), and k-nearest neighbor (kNN) were used as classification algorithms. After the models' training, ten-fold stratified cross-validation was performed. The models were evaluated using the area under the curve (AUC), classification accuracy, and precision.

Results: A total of 36 patients were included in the analysis, 26 GBM and 10 SBM. Models were built using a total of 812 ultrasound images, 435 of B-mode, 265 (60.92%) corresponded to GBM and 170 (39.8%) to metastases. In addition, 377 elastograms, 232 (61.54%) GBM and 145 (38.46%) metastases were analyzed. For B-mode, AUC and accuracy values of the classification algorithms ranged from 0.790 to 0.943 and from 72 to 89%, respectively. For elastography, AUC and accuracy values ranged from 0.847 to 0.985 and from 79% to 95%, respectively.

Conclusion: Automated processing of ultrasound images through deep learning can generate high-precision classification algorithms that differentiate glioblastomas from metastases using intraoperative ultrasound. The best performance regarding AUC was achieved by the elastography-based model supporting the additional diagnostic value that this technique provides.

Keywords: brain tumor; convolutional neural network; deep learning; elastography; intraoperative ultrasound.

Figure 1

Example of intraoperative ultrasound images. 65-year-old man with a right frontal glioblastoma. (A) Elastogram showing the difference in consistency between the tumor and the peritumoral region (green - red) from the rest of the healthy parenchyma (blue). In the right-lower part of the image, a graphic representation of the external compression waves is observed. (B) Simultaneous image in B-mode.

Figure 2

Intraoperative ultrasound images pre-processing. Left: original images of (A) elastogram and (B) B-mode. Right: Final image available for automatic analysis.

Figure 3

Schematic representation of Inception v3 architecture (adapted from GoogLeNet) and the workflow used in the transfer learning process via convolutional neural network and classification algorithms.

Figure 4

Illustrative cases of the use of intraoperative ultrasound. (A) Axial T1 weighted post-contrast (T1WC) image of a 50-year-old man with a right temporal glioblastoma. (B) Elastogram (left) and B-mode (right). It is a soft tumor with small cystic regions and a peritumoral region of low stiffness compared to the healthy parenchyma. (C) Axial T1WC image of a 70-year-old woman with a left occipital glioblastoma. (D) The elastogram shows a cystic/necrotic lesion with a nodular component of intermediate consistency and a relatively soft peritumoral region. (E) Coronal T1WC image of a 45-year-old man with a right parietal lung metastasis. (F) The elastography image shows a solid/cystic lesion with a soft nodular component and a stiffer peritumoral region. (G) Axial T1WC image of a 52-year-old man with no history of systemic cancer with a left parieto-occipital metastasis. (H) The elastogram shows a large cystic lesion with a small hard region and a peritumoral region of similar consistency.

Figure 5

Flow chart of patient and ultrasound image selection process.

Figure 6

Graphical representation of the clusters generated from the distance matrices after analyzing images in (A) B-mode and (B) Elastography. Left: dendrograms of the top two clusters. Right: Bar graph of the probabilities of being assigned to each cluster of glioblastomas (blue) and metastases (red).

Figure 7

Representation of classifier performance using the ROC (Receiver Operating Characteristics) curve for (A) B-mode and (B) Elastography. The best results were obtained by the Support Vector Machine (SVM) and K-Nearest Neighbor (k-NN) algorithms.