A Machine Learning Model to Predict Hepatocellular Carcinoma Response to Transcatheter Arterial Chemoembolization

Abstract

Purpose

To evaluate a fully automated machine learning algorithm that uses pretherapeutic quantitative CT image features and clinical factors to predict hepatocellular carcinoma (HCC) response to transcatheter arterial chemoembolization (TACE).

Materials and Methods

Outcome information from 105 patients receiving first-line treatment with TACE was evaluated retrospectively. The primary clinical endpoint was time to progression (TTP) based on follow-up CT radiologic criteria (modified Response Evaluation Criteria in Solid Tumors). A 14-week cutoff was used to classify patients as TACE-susceptible (TTP ≥ 14 weeks) or TACE-refractory (TTP < 14 weeks). Response to TACE was predicted using a random forest classifier with the Barcelona Clinic Liver Cancer (BCLC) stage and quantitative image features as input, as well as the BCLC stage alone as a control.

Results

The model’s response prediction accuracy rate was 74.2% (95% confidence interval [CI]: 64%, 82%) using a combination of the BCLC stage plus quantitative image features versus 62.9% (95% CI: 52%, 72%) using the BCLC stage alone. Shape image features of the tumor and background liver were the dominant features correlated to the TTP as selected by the Boruta method and were used to predict the outcome.

Conclusion

This preliminary study demonstrated that quantitative image features obtained prior to therapy can improve the accuracy of predicting response of HCC to TACE. This approach is likely to provide useful information for aiding in selection of patients with HCC for TACE.

© RSNA, 2019

See also commentary by Chapiro and Duncan in this issue.

Summary

Prediction of hepatocellular carcinoma response to transcatheter arterial chemoembolization (TACE) using quantitative imaging and clinical measurements of pretreatment lesions is a potentially useful clinical tool that can assist in patient selection for TACE.

Key Points

• ■ Response prediction after initial transcatheter arterial chemoembolization is important for treatment planning.

• ■ Image features on baseline scans combined with clinical staging offer best prediction accuracy.

• ■ Application of machine learning in segmenting tumors and choosing predictive features saves considerable time and effort.

Introduction

Hepatocellular carcinoma (HCC) is the most common primary hepatic malignancy worldwide (85%–90% of primary liver cancers) and the fastest growing cause of cancer-related deaths in the United States, usually occurring in a cirrhotic liver (1–3). It is the second most frequent cause of cancer-related deaths worldwide, causing approximately 750 000 deaths in 2012 (1), with an annual incidence of more than 748 000 newly diagnosed cases globally per year (4). It is also the sixth most common malignancy overall (5) and the fifth most common cancer in men (6). Unfortunately, despite close HCC surveillance in high-risk patients, most diagnoses are intermediate-stage disease to advanced disease (7,8). As a result, curative therapies (surgical resection or liver transplantation) are not available to more than 80% of patients with HCC. Thus, treatment decisions are particularly difficult in this patient population, which has a median survival of 16 months for intermediate HCC and 6–8 months for patients with advanced HCC (9). However, multiple therapeutic options are available for unresectable HCC, including local radiofrequency ablation, embolization, and systemic treatment with sorafenib (10–12).

Local-regional treatment using transcatheter arterial chemoembolization (TACE) selectively delivers high-concentration chemotherapeutics (eg, doxorubicin) to targeted tumors, usually as a stand-alone therapy. TACE takes advantage of the fact that HCCs primarily receive their blood supply from the hepatic artery, whereas the liver parenchyma receives most of its blood supply from the portal vein (13). Thus, TACE can deliver highly concentrated therapy to targeted lesions while sparing the surrounding hepatic tissue (7). However, TACE has multiple adverse effects, including upper quadrant pain, nausea, fatigue, fever, ileus, and elevation of liver enzyme levels. More serious complications may also occur, including liver failure, gastroduodenal ulceration, kidney failure, and death (14–16). Moreover, because any therapeutic approach has a substantial physical, emotional, and financial effect on the patient, reliable methods of predicting the response of HCC to TACE early on during the treatment course are needed. Previous studies suggest that TACE provides a survival advantage for some patients compared with supportive care alone (17). However, predicting which patients with HCC will respond to TACE has proven to be extremely difficult. For example, researchers found that up to 60% of patients with HCC who underwent TACE did not benefit from it despite undergoing multiple sessions (18). Thus, patient selection is essential for effective, safe TACE.

Current patient selection guidelines for TACE are based on the Barcelona Clinic Liver Cancer (BCLC) staging system. This system considers patient performance status, severity of underlying liver disease (Child-Pugh score), tumor size, number of tumors, vascular invasion, and metastasis. According to the BCLC system, patients with intermediate (stage B) HCC are the preferred candidates for TACE. The BCLC staging system has been used globally and has the endorsement of both the American Association for the Study of Liver Diseases and the European Association for the Study of the Liver (9,19). However, the recently proposed Hong Kong Liver Cancer (HKLC) classification showed improved discriminatory ability over BCLC in Asian cohorts (20,21). Sohn et al compared the performances of the five-stage HKLC (HKLC-5), which is essentially a compressed nine-stage HKLC (HKLC-9), with BCLC staging systems in a single United States center cohort of 881 patients and concluded that the HKLC-5 staging system outperformed the BCLC staging system in predicting survival times of patients; this might be a future direction worth exploring (22). Recently, imaging texture analysis using pretherapeutic dynamic CT has drawn interest as a potential predictor of response of HCC to TACE (23). A recent study demonstrated that tumor size of up to 5 cm and single nodularity (single lesion) were predictive of complete response to TACE, whereas multinodularity (multiple nodules within the same lesion) and larger tumor size were associated with recurrence (24). However, tumor size and nodularity are only two of numerous other quantitative image features, and researchers are still working to characterize which image features best correlate to TACE response. Patients with intermediate-stage to advanced HCC would greatly benefit from patient-specific computer models that can identify those persons with a high probability of having response to TACE with preserved liver function. In the present study, we sought to develop predictive models based on image features generated from automated segmentation to identify patients with HCC who would benefit best from TACE prior to the therapy by using quantitative CT image features and clinical prognostic scores.

Materials and Methods

Study Cohort

This retrospective, single-institution study was approved by the MD Anderson institutional review board. It included 105 patients with HCC treated at MD Anderson from November 2002 to June 2012 (Fig 1). The inclusion criteria were TACE as the sole first-line or initial bridging therapy and availability of multiphasic contrast material–enhanced CT images obtained at baseline with no image artifacts (eg, surgical clips). On average, baseline CT was performed 3 weeks before the first session of TACE (range, 1–12 weeks). Of these patients, 68 were men (mean age, 70.5 years [age range, 55–86 years]), and 37 were women (mean age, 74.5 years [age range, 56–93 years]). Eleven, 24, 67, and three patients had BCLC stage A, B, C, and D HCC, respectively. Patients undergoing TACE were administered one of the following chemotherapy regimens: (a) doxorubicin in 20- to 100-mg drug-eluting beads LC Beads (DEBDOX, 45 lesions; BTG International, London, England) or (b) cisplatin, doxorubicin, and mitomycin C (100, 50, and 10 mg, respectively; 60 lesions). The Table shows the patients’ demographic data and clinical profiles.

Figure 1:

Patient cohort selection. HCC = hepatocellular carcinoma, TACE = transcatheter arterial chemoembolization.

Baseline Patient Characteristics

Note.—BCLC = Barcelona Clinic Liver Cancer.

* Mean age was 68 years.

^† Mean α-fetoprotein level was 9551.5 ng/mL.

CT Technique

All patients underwent multiphasic contrast-enhanced CT of the abdomen on four–, 16–, or 64–detector row CT scanners (LightSpeed; GE Healthcare, Waukesha, Wis). The liver protocol was used in all CT studies. The injection rate was 3–5 mL/sec, and the image reconstruction thickness was 2.5–5.0 mm. In total, 105 CT image volumes from the portal venous phases were examined. Scanning during the portal venous phase was performed 60 seconds after peak enhancement of the aorta from contrast material injection. All Digital Imaging and Communications in Medicine images were converted to Neuroimaging Informatics Technology Initiative (NIfTI) for mat to preserve the orientation information and pixel spacing for data processing.

Time to Progression and Response to TACE

The 105 HCC lesions were monitored longitudinally with follow-up CT. Each lesion was monitored for progression based on radiology reports. The time to progression (TTP) was defined as the number of days from the initiation of TACE to when the HCC exhibited radiologic evidence of progression according to the modified Response Evaluation Criteria in Solid Tumors. Lesions were censored if they had not progressed by the time the study was initiated, or if patients were lost to follow-up, died before their target HCCs progressed, or were switched to a different treatment, such as sorafenib, radiation therapy, and surgery. Lesions were divided into TACE-susceptible and TACE-refractory groups based on the radiologic TTP. TACE-susceptible patients had no appreciable radiologic progression at follow-up CT, and TACE-refractory patients showed evidence of radiologic progression at follow-up CT (not necessarily nonresponders because they might respond to repeat chemoembolization session). The TTP cutoff used to stratify response to the first TACE session was 14 weeks; this interval was selected based on the expected date for repeat chemoembolization that was around 14 weeks on average after the first TACE session in our patient cohort. All of our patient cohort underwent follow-up CT within the 14-week window prior to undergoing the second TACE session.

Segmentation Training Data

Our overall approach (Fig 2) used image features extracted from a segmentation of the tumor and background liver to predict TACE response. Neural network–based segmentation models were used to automatically segment the tumor and background liver. Training data for our approach consisted of two distinct datasets. The first dataset consisted of the axial CT images and the corresponding liver segmentations from the Medical Image Computing and Computer Assisted Intervention Society Liver Tumor Segmentation (or LiTS) challenge (25). This dataset consists of 130 manually labeled CT image and segmentation pairs and is publicly available. Manual segmentation (contouring) is not to be confused with Couinaud segmentation of the liver. The public LiTS data were used to train a neural network model to segment the liver within the image.

Figure 2:

The response prediction pipeline. The study cohort consisted of 105 patients with hepatocellular carcinoma (HCC). Each patient underwent transcatheter arterial chemoembolization before CT with subsequent follow-up CT. An automatic segmentation algorithm was created based on a neural network classification model with training performed by using manually segmented images. Image features for all HCCs were extracted and combined with the Barcelona Clinic Liver Cancer (BCLC) score to further predict their response to TACE. CNN = convolutional neural network.

The second dataset consisted of all 105 lesions in our cohort; each three-dimensional lesion was manually segmented from the surrounding tissue, section by section by three radiology residents (A. Morshid, A.M.K., and M.M.E.) and reviewed by a board-certified radiologist with 20 years of experience in abdominal imaging (K.M.E.). The standard deviation of the Dice similarity coefficients (DSCs) between each reader was used to quantify potential labeling bias. Segmentations were performed using the Amira software package (Thermo Fisher Scientific, Waltham, Mass). The abdominal window (for soft tissue: width, 400 HU; level, 50 HU) and liver window (for hepatic lesions: width, 150 HU; level, 30 HU) used in Amira were set to match the picture archiving and communication system viewer used for clinical diagnosis (26). Amira is a commercial package with multiple user groups and has been validated for imaging segmentation studies in the literature (27). Tissues were manually segmented, on the venous phase, into two divisions: (a) background liver parenchyma without disease and (b) tumor including both enhancing tumor portion and nonenhancing tumor. These manual segmentations provided the training data used to develop a neural network classifier for segmentation of tumor from the background liver.

Segmentation Model

Two convolutional neural network (CNN) models (CNN1 and CNN2) were constructed to segment the tumor from the background liver. CNN1 was trained for binary segmentation of liver tissue using an axial portal venous phase CT image as input. CNN2 was trained to segment tumor using an axial portal venous phase CT image and the output of CNN1 as simultaneous inputs. Network architecture follows the U-Net architecture described in Vorontsov et al and Chlebus et al (28,29). Briefly, each CNN is constructed from a composition of convolution and downsampling operations that extract features along a contracting path. Similarly, an expanding path consists of convolution and upsampling operations with long skip connections to integrate features from the corresponding downsampling operations. Four resolution levels are used. Each convolution operation uses a 3 × 3 kernel size and is followed by a batch normalization and a rectified linear unit activation function. A dropout (P =.5) was used before each convolution in the upsampling path. The corresponding code for both models is found in our github repository (https://github.com/fuentesdt/livermask).

Both CNN1 and CNN2 were trained directly from the available data. Both were two-dimensional models and applied section-by-section to the axial images. No pretraining was used. CNN1 was trained on all 130 datasets of the LiTS challenge. The output of CNN1 was postprocessed to use the largest three-dimensional connected component. CNN2 was trained to the 105 tumors of our manually labeled dataset. No postprocessing was applied to CNN2. The accuracy of both models was evaluated with fivefold cross-validation on the respective datasets. The DSC was used to quantify the overlap of the model-generated segmentation with the manual training data.

Survival Prediction Algorithm

Image texture features were extracted from each label of the segmented images representing the background liver and tumor. Shape and texture features from each label were obtained using the pyradiomics package (30). Five feature classes were considered and included the shape features, first order features, gray-level co-occurrence matrix (GLCM) features, gray-level size-zone matrix (GLSZM) features, and gray-level run-length matrix (GLRLM) features. These feature classes are enabled in the pyradiomics packages using the “shape,” “firstorder,” “glcm,” “glrlm,” and “glszm” options, respectively. Mathematical definitions of each specific feature are provided within the supplemental methods of van Griethuysen et al (30).

Because of a large number of imaging features generated, variable reduction was performed on the shape and texture features. Features with a pairwise correlation greater than 0.8 were eliminated to reduce redundant model inputs. The Boruta feature selection method (31) was applied sequentially as a second variable reduction technique to identify predictive image features that were the most useful for TACE response prediction (31). This selection method has been shown to be effective in other high-dimension classification schemes and specifically in genetic studies of the microbiome (32,33).

A random forest binary classifier was used for machine learning prediction of TACE-susceptible and TACE-refractory using the BCLC stage alone and the BCLC stage plus predictive image features as input to the classifier. To avoid model overfitting, prediction accuracy was evaluated using leave-one-out cross-validation. As the name implies, the classifier was first trained on all lesions except one. The trained model was then used to predict the response status of the remaining lesion (susceptible vs refractory). This was repeated for each lesion, and the percent accuracy for predictions was calculated by dividing the number of times the algorithm correctly predicted the response by the total number of trials. The entire response prediction process pipeline is shown in Figure 2. The true-positive rate and false-positive rate were also calculated for each classifier probability threshold and the corresponding area under the receiver operating curve was calculated.

For completeness, two sets of image features were considered in the response predictions. One set of image features was obtained from the CNN2 tumor segmentation output. Importantly, the validation set of the fivefold cross-validation was used to ensure a representative automatic segmentation was analyzed. As a control, a second set of image features was obtained from one of the manually labeled tumor datasets.

Results

Segmentation Accuracy

Representative segmentation of HCCs are shown in Figure 3. Manual segmentations are shown as “truth” and serve as both the reference standard and reference for training our neural network classification model for the tumor volumes. Automated represents tissue classification using the neural network model; here the model is applied to a validation set image that is independent of the training set. The mean ± standard deviation of three reference annotations by humans of the computed DSC was 0.71 ± 0.21 for the tumor volume. Neural network performance correlated with tumor volume. For tumors greater than 10 cm in diameter, the mean ± standard deviation DSC with manually labeled data was 0.67 ± 0.20. For tumors less than 10 cm in diameter, the mean DSC with manually labeled data was 0.33 ± 0.24. Correlation analysis between the manual repeat tumor volume labels and the neural network model accuracy is shown in Figure 4. As measured by the DSC, neural network accuracy was correlated with the DSC between two independent observers (r = 0.755), and inversely correlated with the variability in DSC between all three observers (r = −0.536).

Figure 3:

Images show segmentation of hepatocellular carcinomas using manual and automated techniques. The first column (A, D, and G) shows native livers without any overlying masks. The second column (B, E, and H) shows liver and tumor masks manually segmented. Red = normal liver; green = tumor. The third column (C, F, and I) shows the neural network fully automated segmentation masks.

Figure 4a:

Correlation analysis between the (a) Dice similarity coefficients of the neural network versus the independent observers and between (b) the variability in observer-generated tumor volume labels and the Dice similarity coefficients of the neural network. The contrast-to-noise ratio (CNR) is also shown corresponding to the color legend in a. The computed correlation coefficients (r) as well as P values estimating statistically significant correlation changes from zero are provided with each graph. DSC = Dice similarity coefficient; STD = standard deviation.

Figure 4b:

Correlation analysis between the (a) Dice similarity coefficients of the neural network versus the independent observers and between (b) the variability in observer-generated tumor volume labels and the Dice similarity coefficients of the neural network. The contrast-to-noise ratio (CNR) is also shown corresponding to the color legend in a. The computed correlation coefficients (r) as well as P values estimating statistically significant correlation changes from zero are provided with each graph. DSC = Dice similarity coefficient; STD = standard deviation.

Predicted TACE Response Outcomes and Prediction Accuracy

Upon applying the random forest binary classification model, we found that using the BCLC stage as the sole input in the classifier resulted in a prediction accuracy rate of 62.9% (95% confidence interval [CI]: 52%, 72%) regarding TACE-susceptible versus TACE-refractory. Upon addition of predictive image features extracted from the neural network segmentation to BCLC stage, the prediction accuracy rate increased to 74.2% (95% CI: 64%, 82%). The specific image features added from our algorithmic approach were the tumor volume, the maximum two-dimensional axial diameter of the background liver, the small area low gray level emphasis within the background liver, the maximal correlation coefficient within the background liver, and the long run high gray level emphasis within the tumor. As a reference, addition of these image features extracted from the manual segmentations to BCLC stage resulted in a prediction accuracy of 67.6% (95% CI: 57%, 76%). The corresponding receiver operating characteristic curve for our classifier is shown in Figure 5. The area under the receiver operating characteristic curve was 0.73.

Figure 5:

Receiver operating characteristic analysis of our model predictions of responders and nonresponders. The true-positive rate (sensitivity) is plotted against the false-positive rate (1 − specificity). The curve represents the test set prediction of the leave-one-out cross-validation. Gray area = 95% confidence interval.

Discussion

The goal of this study was to determine whether baseline quantitative CT image features of pretherapeutic HCC in combination with clinical assessment scores can be used to improve response prediction of this cancer to TACE. A similar approach was recently explored by Abajian et al using MRI and multiple clinical features (34). The results when using a logistic regression classifier were 72% and when using a random forest classifier were 66%. Our approach differed in developing an in-house automated pipeline using derived quantitative image features, such as tumor volume, and the BCLC clinical prognostic score only to predict response. For patients with multiple therapies, only the initial TACE treatment was considered to avoid confounding treatment factors. Our results demonstrated that use of the BCLC stage alone resulted in a response prediction accuracy rate of 62.9%. The accuracy rate improved to 74.2% when including both the BCLC stage and predictive image features from our neural network segmentations into our classifier. The same image features from the manual segmentation of the tumor volumes achieved lower accuracy. This was likely the result of combination of the variability in the manual image segmentation and errors in the neural network segmentation observed in Figure 4. Neural network segmentation does not always agree with manual segmentation. However, the consistency in the machine decision criteria leads to consistent image features that achieve the best performance during cross-validation.

BCLC staging is currently used to stratify risk for patients with HCC who are planning to undergo TACE (35). Our results indicate that combining clinical data with relevant CT image features can improve prognostication. Although it has been shown that tumor volume, and specifically necrotic tumor volume, is predictive of response (18,30), our work demonstrates that other components on CT images, such as the size of the background liver and quantitative image features, are also key factors. Other textural features proposed to be prognostic for HCC progression include tumor hypervascularity and tumor margin growth (31–33). Future efforts can work to incorporate these metrics into our existing classification framework as well as systematically evaluate the effects of image preprocessing on outcome prediction accuracy.

One of the limitations in this study was that our prediction model is general and was applied to multiple TACE chemotherapy regimens including drug-eluting bead TACE and conventional TACE. Triple-drug TACE was found to have no survival difference from single-drug TACE using doxorubicin-eluting beads (36). Drug-eluting bead TACE was also not found to have superior survival benefit over conventional TACE in recent literature (37,38); however, multiple studies expect the objective response rates to be improved for drug-eluting bead TACE (39–41). Additional model prediction accuracy of TTP may be possible by including the TACE delivery parameters in the model predictions. A future direction would also include comparing outcomes between drug-eluting bead TACE and conventional TACE when analyzed separately. In addition, 10 patients did not have any progression at follow-up CT prior to their death, in which tumor progression could have been present but not measured.

The ideal TACE candidate is a patient with unresectable HCC and BCLC stage B disease. One of the limitations of this study was the oversized sample of patients with BCLC stage C and D disease (n = 70) at the time of the first TACE session. These patients have higher disease burden than patients with BCLC stage B disease and would therefore mostly be TACE-refractory and require switching to tyrosine kinase inhibitors. Resolution of the CT scanners was a limiting factor for detecting progression in this study. Any microscopic progression undiscernible to a radiologist was undetected. In addition, standard-of-care imaging protocols were used. Under these protocols, patients underwent imaging with variability in the resolution of each CT scan. The automated segmentation training algorithm was also limited by the variability in the manual segmentation tumor volumes. We attempted to quantify this variability by having three radiologists cross-validate the work, measure the variance statistics, as well as measure agreement with human error (Fig 4). Similar to the study by Vorontsov et al (28), we report accuracy as a function of tumor volume. In general, neural network segmentation performs better on larger lesions. Also, the fact that HCC is usually found in cirrhotic livers that have a heterogeneous parenchyma compared with healthy liver could have affected the automated algorithm’s ability to distinguish between tumor and background liver tissue. The delayed phase of the contrast-enhanced scan is acquired at 3–5 minutes after contrast material injection and was not used in this study. While in the clinical setting the delayed phase is used to accurately define the tumor boundaries, breathing motion was seen to cause significant registration errors. Future efforts will incorporate delayed phase information into the segmentation model and evaluate overall effect on response prediction accuracy on a larger BCLC stage B patient population. Another limitation of the current work was that it relied on the available radiology reports from the time the scans were read to assess for disease progression. The reading radiologists were all fellowship-trained, qualified, practicing radiologists in a large academic center; thus, this inherently could introduce interreader variability in assessing progression. A future work is currently underway to reassess follow-up CTs by the same group of radiologists to avoid this limitation.

From a public health perspective, this predictive model may help lessen the burden placed on patients and the health care system when TACE is administered to refractory patients, and medical resources are not optimally allocated. Accurately prognosticating the outcome of TACE before starting the course of therapy would save time, money, and opportunity cost for both patients and physicians. Moreover, if TACE results in a partial response, then repetition of the procedure may be indicated. Post-TACE TTP was found in the literature to correlate with overall survival (42). Lesions with a TTP of less than 5 months were branded as TACE-refractory and would not benefit from further TACE sessions. Lesions with a TTP equal to or greater than 5 months were branded as TACE-susceptible and would benefit from further TACE sessions (42–44). However, a delay in determining whether the initial TACE procedure is successful may limit the benefits of repeated TACE. Therefore, methods that can accurately identify which patients would benefit from TACE and the extent of disease progression after therapy are extremely useful.

A future direction to consider would be identification of genetic and biochemical features of HCCs that can be incorporated into the prediction schema. Two studies have suggested that expression of CD34-positive and vascular endothelial growth factor–negative tumor markers are associated with TACE resistance, possibly the result of adaptive mechanisms making tumors resistant to hypoxia (18,45). However, use of these two pathologic markers would require biopsy analysis prior to TACE. Another study identified a panel of pre-TACE serum protein markers (including α-fetoprotein, protein-induced by vitamin K absence of antagonist-II, leucine-rich α-2-glycoprotein 1, serum amyloid P component, butyrylcholinesterase, complement component 7, and ficolin-3) that can help predict TACE response of HCC (46). These factors can be investigated further to improve response prediction accuracy and selection of patients with HCC for TACE. We found that prediction of HCC response to TACE using quantitative imaging and clinical measurements of pretreatment lesions is a potentially useful clinical tool that can assist in patient selection for TACE.