Development of deep neural network for individualized hepatobiliary toxicity prediction after liver SBRT

Abstract

Background: Accurate prediction of radiation toxicity of healthy organs-at-risks (OARs) critically determines the radiation therapy (RT) success. The existing dose-volume histogram-based metric may grossly under/overestimate the therapeutic toxicity after 27% in liver RT and 50% in head-and-neck RT. We propose the novel paradigm for toxicity prediction by leveraging the enormous potential of deep learning and go beyond the existing dose/volume histograms.

Experimental design: We employed a database of 125 liver stereotactic body RT (SBRT) cases with follow-up data to train deep learning-based toxicity predictor. Convolutional neural networks (CNNs) were applied to discover the consistent patterns in 3D dose plans associated with toxicities. To enhance the predicting power, we first pretrain the CNNs with transfer learning from 3D CT images of 2644 human organs. CNNs were then trained on liver SBRT cases. Furthermore, nondosimetric pretreatment features, such as patients' demographics, underlying liver diseases, liver-directed therapies, were inputted into the fully connected neural network for more comprehensive prediction. The saliency maps of CNNs were used to estimate the toxicity risks associated with irradiation of anatomical regions of specific OARs. In addition, we applied machine learning solutions to map numerical pretreatment features with hepatobiliary toxicity manifestation.

Results: Among 125 liver SBRT patients, 58 were treated for liver metastases, 36 for hepatocellular carcinoma, 27 for cholangiocarcinoma, and 4 for other histologies. We observed that CNN we able to achieve accurate hepatobiliary toxicity prediction with the AUC of 0.79, whereas combining CNN for 3D dose plan analysis and fully connected neural networks for numerical feature analysis resulted in AUC of 0.85. Deep learning produces almost two times fewer false-positive toxicity predictions in comparison to DVH-based predictions, when the number of false negatives, i.e., missed toxicities, was minimized. The CNN saliency maps automatically estimated the toxicity risks for portal vein (PV) regions. We discovered that irradiation of the proximal portal vein is associated with two times higher toxicity risks (risk score: 0.66) that irradiation of the left portal vein (risk score: 0.31).

Conclusions: The framework offers clinically accurate tools for hepatobiliary toxicity prediction and automatic identification of anatomical regions that are critical to spare during SBRT.

Keywords: SBRT; convolutional neural networks; liver cancer; toxicity prediction.

© 2018 American Association of Physicists in Medicine.

Figures

Figure 1

Examples of portal vein (PV) with superimposed SBRT doses. (a) The portal vein with the corresponding CT image and isodose curves. (b) A 3D dose plan delivered to the central hepatobiliary tract (15 mm expansion of the portal vein), which was analyzed using deep learning for hepatobiliary toxicity prediction. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 2

Alignment of 3D dose plans delivered to portal vein according to anatomical segments of the portal vein. Rigid registration is used to avoid deformations of dose plans. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 3

The convolution neural network architecture developed for hepatobiliary toxicity prediction.

Figure 4

ROC curves for the proposed deep dose analysis, alternative machine learning‐based toxicity predictors and existing DVH‐based predictors. (a) All SVM, RFs, FcNNs, CNNs and combined CNN and FcNN predictors achieved the Area under receiver operating characteristic curve (AUC) > 0.7. (b) The combined predictor with AUC of 0.859 outperforms the existing DVH‐based predictors VBED 1030, VBED 1040, and VBED 1030 U VBED 1040. The 95% confidence interval (95% CI) was reported for all AUC values. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 5

Performance of the proposed deep dose analysis using the number of false‐positive and false‐negative predictions. For example, when all SBRTs resulted in toxicities were identified (false negatives = 0), the random forest, fully connected, and convolutional neural networks make 18, 43, and 20 false‐positive predictions. [Color figure can be viewed at wileyonlinelibrary.com]

Figure 6

Maps of crucial portal vein (PV) subregions irradiation of which correlates with hepatobiliary (HB) toxicity manifestation computed for five randomly selected patients. Among four PV anatomical regions, irradiation of the central bifurcation and proximal PV could most likely cause HB toxicity with the average risk scores of 0.680 and 0.656, while irradiation of the left PV branch could least likely cause HB toxicity with the average risk score of 0.311. [Color figure can be viewed at wileyonlinelibrary.com]