A Lesion-Based Response Prediction Model Using Pretherapy PET/CT Image Features for Y90 Radioembolization to Hepatic Malignancies

Abstract

We present a probabilistic approach to identify patients with primary and secondary hepatic malignancies as responders or nonresponders to yttrium-90 radioembolization therapy. Recent advances in computer-aided detection have decreased false-negative and false-positive rates of perceived abnormalities; however, there is limited research in using similar concepts to predict treatment response. Our approach is driven by the goal of precision medicine to determine pretherapy fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography imaging parameters to facilitate the identification of patients who would benefit most from yttrium-90 radioembolization therapy, while avoiding complex and costly procedures for those who would not. Our algorithm seeks to predict a patient’s response by discovering common co-occurring image patterns in the lesions of baseline fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography scans by extracting invariant shape and texture features. The extracted imaging features were represented as a distribution of each subject based on the bag-of-feature paradigm. The distribution was applied in a multinomial naive Bayes classifier to predict whether a patient would be a responder or nonresponder to yttrium-90 radioembolization therapy based on the imaging features of a pretherapy fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography scan. Comprehensive published criteria were used to determine lesion-based clinical treatment response based on fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography imaging findings. Our results show that the model is able to predict a patient with liver cancer as a responder or nonresponder to yttrium-90 radioembolization therapy with a sensitivity of 0.791 using extracted invariant imaging features from the pretherapy fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography test. The sensitivity increased to 0.821 when combining extracted invariant image features with variable features of tumor volume.

Introduction

Yttrium-90 radioembolization (Y90-RE) is recommended for unresectable, chemorefractory liver-dominant primary or metastatic hepatic disease with a life expectancy of 3 months or longer.¹ Accumulated studies demonstrate that Y90-RE can improve the overall outcome of disease progression, from being unresectable to resectable or convert incurable disease into transplantable and potentially curable in patients with colorectal carcinoma liver metastasis,² neuroendocrine liver metastasis,³ primary hepatocellular carcinoma,⁴ or intrahepatic cholangiocarcinoma.⁵ If a patient, however, is unresponsive to the Y90-RE, the expensive and technically demanding treatment will result in unnecessary risk and cost to the patient.

Imaging assessment of the response to Y90-RE is pivotal for patient management. As patients who are referred for Y90-RE generally have advanced, unresectable, and chemorefractory liver malignancies, fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography (FDG PET/CT) has been widely used in clinical practice for a preprocedural workup. Fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography can depict tumor metabolic activity as well as certain anatomic features including tumor size and lesion density. European Organization for Research and Treatment of Cancer (EORTC)⁶ criteria and PET Response Criteria in Solid Tumors (PERCIST) use the change in tumor standardized uptake value (SUV) to determine tumor response and subsequently analyze patient outcome.⁷ Computed tomography assessments use size-based revised Response Evaluation Criteria in Solid Tumors (RECIST 1.1)⁸ or tumor necrosis-based Choi criteria.⁹ Recent reports have shown that FDG PET/CT was useful in predicting the overall survival after Y90-RE in patients with FDG-avid liver metastases^10,11 or primary liver cancers,¹² especially cholangiocarcinomas.¹³

The objective of this study is to create a model that predicts patient treatment response from the pretherapy FDG PET/CT scan alone. Since FDG uptake has been found to be correlated with microenvironmental tumor characteristics, such as hypoxia, cell proliferation, and blood flow, we hypothesize baseline FDG PET/CT scans contain unique imaging features that are shared across the patient population.¹⁴ Toward this goal, several studies have evaluated the effectiveness using conventional methods such as baseline maximum standardized uptake value (SUV_max). However, the predictive power is considerably low, thus limiting the SUV_max as a prospective predictive tool for patient treatment options.^15,16

In contrast, imaging features captured using machine learning algorithms and pattern classification techniques trained on extracted texture imaging features from tumor lesions have yielded better results in predicting patient response.¹⁷ However, those techniques are limited by changes in lesion orientation and other transformations. Since no lesion is ever the same, the ideal scenario would be able to discover imaging markers that are independent of the state of the tumor and can provide us with a better understanding of tumor morphology.

To address the above challenge, we built a model that extracts invariant texture and shape imaging features from lesions to identify common imaging features. We assume each lesion is a mixture of these imaging features, and the contribution of a distinct feature can be used to differentiate between responders and nonresponders. A binary classification model is utilized to investigate the accuracy of the model, where we train on a set number of lesions extracted from pretherapy FDG PET/CT and test whether the classifier recognizes the treatment response of a new image. The framework is general and can be applied to classify other lesions and the corresponding treatment response. Our empirical results achieve a sensitivity of 0.791 on a set of invariant imaging features.

Materials and Methods

Patients

This retrospective study has been approved by the institutional review board of University of Illinois Hospital & Health Sciences System. A total of 173 Y90-RE procedures were performed between January 2011 and June 2014 at our institution. Patients who underwent both pre- and posttherapy FDG PET/CT scans with at least 3-month clinical follow-up were included. For patients who received multiple consecutive Y90-RE therapies, we treated each patient’s follow-up scan as a pretherapy scan for the next Y90-RE treatment. There were a total of 12 patients (8 patients with liver metastases from colon cancer and 4 patients with cholangiocarcinoma) with 30 Y90-RE procedures that met our selection criteria (Table 1). Patients with complete and partial response were combined into responders (R), whereas stable and progression of disease were combined into nonresponders (NR). On lesion-based analyses (Table 2), our data set was determined to have 17 responders and 13 nonresponders. Of the 17 responders, 4 cases were from 3 patients with cholangiocarcinoma and the remaining 13 cases were from 8 patients with metastatic colon cancer. Of the 13 nonresponders, 4 cases were from 3 patients with cholangiocarcinoma and the remaining 9 cases were from 7 patients with metastatic colon cancer. Representative FDG PET/CT cases of responder and nonresponder for Y90-RE are shown in Figure 1.

Table 1.

Characteristics of Eligible Patients for This Study.^a

Patient Characteristics	Values
Age, years
Median	62.5
Mean	63
Range	30-77
Sex (number of patients)
Male	6
Female	6
Liver cancer (number of patients)
Cholangiocarcinoma	4
Liver metastatic	8
Body surface area, m2
Median	1.78
Range	1.33-2.53
Tumor volume, mm3
Mean ± standard deviation	54.9 ± 11.8
Range	30.4-91.5

^an = 12.

Table 2.

Combined Imaging Criteria for Assessment of Treatment Response.

Tumor Response	EORTC6 and PERCIST7-Based PET Criteria	Choi Criteria (Choi et al) Based on Tumor Necrosis9	RECIST1.18-Based Size Criteria
Complete response	Resolution of metabolic active lesions	Complete necrosis	Disappearance of all lesions, no new lesions
Partial response	25% and 0.8 U decrease of SUVmax of the most intense lesions	More than 15% decrease in tumor density	30% decrease in sum diameters of the target lesions
Progression disease	Greater than 25% increase in SUVmax or more than 20% increase in extent	Not applicable	20% increase in sum diameter of the target lesions
Stable disease	Increase of less than 25% or decrease of less than 15% of SUVmax or no visible increase in extent	Not applicable	<30% decrease or <20% increase in sum diameter or the target lesions

Abbreviations: EORTC, European Organization for Research and Treatment of Cancer; PET, positron emission tomography; PERCIST, PET Response Criteria in Solid Tumors; RECIST1.1, revised Response Evaluation Criteria in Solid Tumors; SUVmax, maximum standardized uptake value.

Figure 1.

Representative FDG PET/CT cases of responders and nonresponders for Y90-RE. Patient 1 with liver metastases from colon cancer is a responder to Y90-RE ([a1 and a2] pretherapy FDG PET/CT; [b1 and b2] posttherapy FDG PET/CT). (a1) Pretherapy FDG PET Maximum intensity projection (MIP) image and (a2) fusion FDG PET/CT image showing multiple hypermetabolic bilobar liver metastases; (b1) posttherapy FDG PET MIP image and (b2) posttherapy FDG PET/CT image demonstrating resolution of hypermetabolic liver metastases. Patient 2 with liver metastases from colon cancer is a nonresponder to Y90-RE ([c1 and c2] pretherapy FDG PET/CT; [d1 and d2] posttherapy FDG PET/CT). (c1) Pretherapy FDG PET MIP image and (c2) fusion FDG PET/CT image showing multiple hypermetabolic bilobar liver metastases; (d1) posttherapy FDG PET MIP image and (d2) posttherapy FDG PET/CT image demonstrating progression of hypermetabolic liver metastases. Additional multiple new hypermetabolic intraperitoneal metastases also developed (d1, arrows). FDG PET/CT indicates fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography; Y90-RE, yttrium-90 radioembolization.

Patient Treatment

All patients underwent hepatic arteriography with Tc-99 m macroaggregated albumin to detect extrahepatic shunting 7 to 14 days prior to Y90-RE. Flow to extrahepatic organs was corrected by coil embolization of shunting vessels or catheter positioning. None of the patients had greater than 20% lung shunt fraction, lung exposure more than 30 Gy in a single Y90-RE procedure, or more than 50 Gy in multiple Y90-RE sessions.

Yttrium-90 radioembolization procedures were performed according to the published protocols.¹⁸ Briefly, Y90 resin microspheres (SIR-Spheres; Sirtex Medical, Lane Cove, Australia) were delivered into a lobar branch of the hepatic artery supplying the tumors. The Y-90 dose was prescribed based on a published body surface area method.¹ The tumor and total liver volume ratio was calculated using CT.

Patient Imaging

Fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography examinations were performed within 2 weeks before treatment and at ∼4 weeks and 3 months after treatment on a GE Discovery 690 PET/CT scanner (GE Medical Systems, Milwaukee, Wisconsin) using the standard protocol. Patients fasted at least 4 hours before scanning and had a blood glucose level <200 mg/dL at the time of FDG injection. Dedicated PET/CT scans from the skull base to the upper thighs were obtained 60 to 90 minutes after intravenous injection of 10 to 13 mCi of FDG with CT parameters of 120 kV, 120 mA, pitch 0.813, and 16 × 1.5 mm collimation, and PET parameters of 3 minutes/bed position.

Clinical Evaluation of Treatment Responses

The FDG PET/CT images were analyzed retrospectively on a dedicated AW PACS workstation (GE Medical Systems). Lesion-based treatment responses were clinically evaluated according to a combined criteria: PET-based EORTC⁶ and PERCIST⁷ criteria, tumor necrosis-based Choi criteria,⁹ and tumor size-based RECIST1.1⁸ (Table 2), with consensus among physicians.

In conflicting cases, the criterion with the best response was used to determine the patient response. For example, the SUV change within 25% (stable on PERCIST), but decreases in size by 30% (partial response on RECIST), or necrosis by 50% (partial response on necrosis criteria), was considered partial response. Resolution of FDG avidity and/or complete necrosis of treated lesions were considered complete response regardless of changes in lesion size. In addition, occurrence of new lesions in the treated liver lobe was considered progressive disease regardless of changes in SUV, necrosis, or lesion size.

Quantitative Analysis

Three-dimensional lesion extraction

For every FDG PET/CT image slice, the cross-section lesion was separated using the spatial fuzzy C-means clustering algorithm.^19,20 This technique automatically assigned image voxels to a cluster based on their relative distance to each other and the correlation between their intensities. The subsequently optimized clusters based on both spatial and intensity characteristics allowed us to filter out lesions with partial volume effect or voxels of noise.²¹ The 3-dimensional (3D) representation of an extracted lesion was rendered by interpolating image slices along the aligned cross-section directions using cubic spline interpolation (Figure 2). To improve computational efficiency, lesions were stored in a 41 × 41 × 41 cube, as this size was able to incorporate all the various lesion sizes. Although the cube matrix was the same for all lesions, larger lesions would take up more space within the cube than smaller lesions.

Figure 2.

Segmentation of liver lesions. A, Representative 3 FDG-avid liver metastases with different morphologies in a patient with metastatic colon cancer marked as 1, 2, and 3, respectively. B, The region of interests of the 3 liver metastatic tumors are extracted using spatial fuzzy clustering and then used to interpolate into 3D space. C, The 3D rendering of the lesions where imaging features that are extracted quantify shape, texture, and intensity. 3D indicates 3-dimension; FDG, fluorine-18-2-fluoro-2-deoxy-d-glucose.

Image feature extraction techniques

The image feature extraction techniques were chosen based on previous studies to have invariant texture and shape features that were highly informative for characterizing disease-specific lesions.^22,23 These techniques have also found success in 3D shape retrieval. As any metastatic lesion deforms into a 3D object, we use these methods to find the common patterns of the lesions.²⁴ In the computer-aided FDG PET/CT image analyses, we used 3D spherical Gabor filter (3DSG) as the texture feature to probe the spatial interrelationships and arrangement of the basic elements of the tumor,²⁵ 3D Zernike descriptor (3DZD) to extract co-occurring geometric patterns, and wave kernel signature (WKS) for the shape features.²⁶ Details on each algorithm are described as follows.

Three-dimensional spherical Gabor filter

Specifically, a 3DSG filter is the product of a 3D Gaussian and a complex exponential function representing a sinusoidal plane. We created multiple filters by changing the shape (σ) rotation (θ, φ), and central frequency (F) of the Gaussian. This filter therefore responds to some frequency but only in a localized part of the image to obtain different representations of our original image. These sets of new images were obtained by convolving the original image, I(x, y, z), with each filter. The convolved images described the number of voxels occurred within our region of interest during a cycle of periodically repeating intensity variations.²⁷ The parameters used for creating the filters were based on obtaining a set of Gabor filters that could acquire the largest variation, which were determined to be 0.125 cycle/pixel for the highest frequency and 0.03125 for the lowest frequency. Given these criteria, the following parameters that achieve the best performance are:

$$\left[F=\text{π}{2}^{\frac{i+2}{2}}, \text{θ}=k\frac{\text{π}}{8}\text{,ϕ}=j\frac{\text{π}}{8}\right],$$

where i = 2, 3, 4, 5, 6, j = 0, …, 7, and k = 0, … , 7. In addition, we set the shape of the envelope by setting $\text{σ}=\frac{1}{F}$. The resulting filter bank contained 320 3D Gabor kernels (5 × 8 × 8) of size 41 × 41 × 41 that were convolved with the original image of the lesion, I(x, y, z). To obtain rotation and scale invariant features, we calculated the 3D discrete Fourier transform (DFT) of the mean and standard deviations of the convolved images, as the magnitude of the DFT has been shown as shift invariant. Each lesion was then represented with a feature vector of size 1 × 640.

Wave kernel signature

The WKS feature descriptor provides robust analysis to nonisometric perturbations of the surface.²⁸ This is imperative in our work as tumors have no predefined shape and are anisotropic (grow in many directions). Furthermore, WKS is able to differentiate between the small microstructures that formed on the surface of a lesion. The application of WKS was accomplished by first modeling the lesion as a geometric shape bounded by a mesh. The wave equation was then applied to a particle with an energy distribution. Given the energy distribution, the probability of a particle being at vertex (x, y, z) was calculated. We adopted parameters with the desired amount of signatures of 100 and variance of 7 as in the literature,²⁸ resulting in a feature vector of dimension 1 × 100.

Three-dimensional Zernike descriptor

The 3DZD is an extension of geometric moments that are used to describe an image based on its area or image orientation. The 3DZD was developed using the Zernike polynomials to describe geometrical patterns while remaining invariant to rotation and scale. The advantage of 3DZD is its ability to describe nonspherical-like shapes into which tumor shapes tend to deform.²⁴ The 3DZD was achieved by first converting the original image Cartesian coordinates, I(x, y, z), into a set of spherical coordinates, I(r, θ, φ), and then evaluating the Zernike polynomial at that point with parameters n, m, and l. The specific 3DZD of order nml was then obtained by integrating the products of Zernike polynomial and the original image. We use the reported optimal parameters²⁹ of n, m, and l as 20, 14, and −14, respectively, with a resulting feature vector size of 1 × 122.

Spatially sensitive bag of features and visual words

Given the large size of the feature vectors in comparison with the size of our data set and the lack of interpretability of the features, we used spatially sensitive bag of features (SSBoF) to summarize the features into meaningful information. The SSBoF representation is an adaptation of the bag-of-words scheme used for text categorization and text retrieval. The key idea is the construction of a code book—that is, a visual vocabulary—in which the most representative patterns are codified as code words or visual words. Then, the image representation is generated through a simple frequency analysis of each code word inside the image.³⁰ The advantage of SSBoF over the generic bag-of-features algorithm is that it also accounts for the relationship between similar code words. As shapes have similar geometric features, using SSBoF is more advantageous as it creates a set of more unique visual words.

As a result of SSBoF, each lesion is represented by a histogram, where each bin signifies a visual word and the counts indicate how often a specific visual word occurs in the lesion. We created an SSBoF vector for each of the original feature descriptors (3DSG, WKS, and 3DZD) by finding the codebook sizes with optimal model performance. We also created an SSBoF vector that combines all the feature descriptors to determine if performance can be improved. To remove any potential bias caused by our assumption that a patient’s follow-up FDG PET/CT scan can act as a pretherapy scan for subsequent Y90-RE treatment, we added a visual word to all histograms denoting whether a patient had previously received Y90-RE therapy.

Multinomial naive Bayes classifier and statistical analyses

We used a multinomial naive Bayes (MNB) classifier to determine whether patients are responders or nonresponders.³¹ The MNB classifier is a specific instance of a naive Bayes classifier that uses a multinomial distribution for each of the features. In our computer-aided FDG PET/CT image analyses, the distribution of the MNB classifier was a histogram of visual words that described each lesion. Our training set size was 22 FDG PET/CT scans, with 13 responder cases and 9 nonresponder cases; our test set was of size 8 with 4 responder cases and 4 nonresponder cases. In order to overcome the small data set size, we used bagging to create multiple training sets by sampling the training set with replacement.³² An MNB classifier was then trained on each new training set, and a patient was classified as a responder or a nonresponder based on averaging the output class probabilities computed by each MNB.

Performance of the classifier was evaluated using a precision/recall characteristic. Precision, or positive predictive value, is the number of correct predicted responders (true positives) divided by the total number of predicted responders (sum of true and false positives). Recall, or sensitivity, is the number of correct predicted responders (true positives) divided by the total number of clinical defined responders (sum of true positives and false negatives). We also used the Kruskal-Wallis test to assess whether there was a significant difference in the performance of the classifier when using different feature descriptors and the Dunn test to compare between individual groups.

Results

The experiments were geared toward decreasing unnecessary procedures and patient risk, so we chose parameters that would minimize the number of false responders, that is, false-positive rate. Given this condition, we found the optimal size of our codebook to be 17, 8, and 11 for 3DSG, WKS, and 3DZD, respectively (P ≤ .05). The precision and recall results are listed in Table 3. The WKS feature descriptor provided the best recall (0.763) and precision (0.833) in comparison with 3DZD (P ≤ .05) and marginal improvement in comparison with 3DSG. As WKS was used to identify anisotropic shape features, we can conclude that there is a correlation between lesion growth pattern and patient treatment response.

Table 3.

Various Feature Groups and the Resulting Recall and Precision.

Features	Recall	Precision
SUVmax (S)	0.485	0.76
Tumor volume (V)	0.583	0.631
3D gray level co-occurrence matrix (C)	0.53	0.77
3D spherical Gabor (G)	0.716	0.795
WKS (W)	0.763	0.833
3D Zernike descriptor (Z)	0.611	0.811

Abbreviations: 3D, 3-dimension; SUV, standardized uptake value; WKS, wave kernel signature.

To test whether the new imaging features enrich lesion analysis, we compared invariant imaging features with SUV_max, 3D gray level co-occurrence matrix (3DGLCM), and tumor volume. We obtained similar receiver operating characteristic curves (Supplemental Figure 1) to that obtained in previous research.^17,33 As shown in Table 3, using tumor volume or SUV_max as a sole predictor resulted a recall of 0.485 and 0.583 and a precision of 0.76 and 0.631, respectively, significantly lower than (P ≤ .05) those of 3DZD, 3DSG, and WKS.

We hypothesized combinations of the various features would improve the results. Using the same conditions as before, we obtained an optimum codebook size of 14 when combining 3DZD, 3DSG, and WKS as 1 descriptor (G + W + Z). The resulting descriptor is also concatenated with SUV_max, 3DGLCM, and tumor volume (Table 4). We omitted the models with different combination features that didn’t show any added performance gains. With the addition of tumor volume as a feature (G + W + Z + V), there was a significant increase (P ≤ .05) in recall (0.821) compared to that of G + W + Z (recall 0.791). Adding the 3DGLCM or SUV_max as a feature did not offer more insight into patient response and decreased predicting power. The difference between the various combinations of features illustrates the value of volume as a distinct global feature to improve the performance.

Table 4.

Various Fused Feature Groups and the Resulting Precision and Recall.

Fused Features	Recall	Precision
G + W + Z	0.791	0.839
G+ W + Z + S	0.758	0.763
G + W + Z + C	0.683	0.717
G + W + Z + V	0.821	0.844

Abbreviations: C, 3D gray level co-occurrence matrix; G, 3D spherical Gabor; S, maximum standardized uptake value; V, tumor volume; W, wave kernel signature; Z, 3D Zernike descriptor.

The analysis of the visual words can provide important information regarding the most distinctive features shared among different patient groups (Figures 3 and 4). This experiment was done using the information of the invariant features. The average visual word histograms for responders and nonresponders were computed using all the histograms obtained from the model. We then assessed the visual word that was distinctive between the 2 patient groups. We found that the 10th bin for 3DSG, fifth bin for 3DZD, and second and seventh bin for WKS were the most pronounced in responders. In the case of nonresponders, the fourth bin for 3DSG, third bin for 3DZD, and first and eighth bin for WKS were found to be most pronounced. These data suggested that there are commonly occurring imaging patterns within the corresponding groups of nonresponders and responders.

Figure 3.

Average visual words for responders (top) and nonresponders (bottom): 10th bin for 3D spherical Gabor filter (3DSG), 5th bin for 3D Zernike descriptor (3DZD), and 2nd and 7th bin for wave kernel signature (WKS) were the most pronounced in responders, whereas the 4th bin for 3DSG, 3rd bin for 3DZD, and 1st and 8th bin for WKS were found to be most pronounced for nonresponders. These visual words can be used to determine patients who are more likely to respond to treatment.

Figure 4.

The most common visual words (left column) for the corresponding largest tumor lesions (right column) in the representative responder (upper row) and nonresponder (lower row). Visual acuity is not enough to distinguish between tumors without further analysis. Using invariant features allows us to determine which tumors will respond to treatment. For example, in the case of 3D spherical Gabor filter, the high occurrence of the 10th bin indicates a likely patient response, as shown in responder 2 (upper row); the high occurrence of the fourth bin describes a tumor that will not respond, as shown in nonresponder 3 (lower row).

Discussion

Having the ability to predict the outcome of a treatment based on pretherapy FDG PET/CT may avoid unnecessary patient risks and expensive, invasive procedures, along with the potential to provide precision treatments. Our study focuses on using pretherapy FDG PET/CT data to establish an Y90-RE response prediction model. Imaging analysis of a patient’s lesion characteristics is important for tumor response evaluation after therapy. Exploration of the commonly used image features such as lesion size, tissue density, and lesion SUV_max on FDG PET/CT in predicting tumor response have found them to be limiting. Meanwhile, computer-aided imaging analyses in evaluation of tumor response have been showing promising results. Tixier et al used local homogeneity as a texture feature to identify patients who would respond to chemoradiotherapy with an area under the curve (AUC) of 0.7 in comparison with the SUV, which had an AUC of 0.59.³³ Similarly, Tan et al showed a tumor with a greater score in homogeneity is more likely to respond to therapy.¹⁷

The preliminary success of using texture features in tumor imaging analyses indicates there is a large amount of information in a radiological image that is not discernible by conventional tools. We demonstrate that a predictive algorithm based on imaging features extracted from pretherapy FDG-PET/CT scans is able to predict patients as responders or nonresponders of Y90-RE with respect to the choice of features. Similar to the study by Tan et al, we showed that SUV_max as an imaging feature could not provide spatial or topological information. Actually, upon analyses of individual image features (Table 3), the commonly used clinical image features, SUV_max and tumor volume, were associated with the lowest predication recall and precision. This is probably due to the fact that these are global features and cannot capture the heterogeneity within a lesion.³⁴ Moreover, we found texture features derived from 3DGLCM also had low prediction power most likely because the feature was susceptible to changes in transform and scale, thus unable to detect subtle differences among tumors. On the contrary, invariant features, in particular WKS, provided the highest recall and precision (P ≤ .05) in predicting a patient’s response to Y90-RE. We conclude that imaging features that are able to describe the heterogeneous nature of a lesion provide richer information. This finding is consistent with the recent reported data on a 3D spatial model of tumor evolution suggesting that cell motility is a key factor in the initial growth of a tumor mass.³⁵

When combining different feature descriptor groups, we found that the best group of features in improving optimal precision and recall is the combination of the 3D invariant features (3DSG, 3DZD, and WKS) and tumor volume (P ≤ .05). As each feature adds a unique element to the classifier, it is understandable that the performance of the feature group is superior to individual features alone. Of particular interest is the feature of tumor volume. When used alone in the predication model, tumor volume had a low predicative power. However, in combination with the invariant imaging features, the performance increased significantly. This is probably due to the fact that invariant images ignore scale when finding patterns among lesions, yet tumor volume is still a critical factor in determining the prognosis of a patient. Therefore, an improved model should include both invariant imaging features and variant features as tumor volume.

We envision that, insofar as tumor heterogeneity is concerned, there is certain tumor homogeneity, or common features, that are related to underlying gene profiling within the tumor and tumor microenvironment. Beyond the potential to stratify patients further with respect to imaging features, we can see in Figures 3 and 4 unique visual words are associated with patient response to Y90-RE. Therefore, it is possible to identify patients who have the potential to respond to Y90-RE based on their pretherapy FDG PET/CT imaging features.

A major limitation of this study is the small number of cases. Although we attempted to overcome the limitation of a small data set by using a bootstrapped model, a larger data set would surely provide a more robust learning schema and could improve performance. Upon validating our model further on a larger data set, we hope to provide a disease- and treatment-specific prediction model based on pretherapy FDG PET/CT image features.

Conclusion

In summary, we developed a model that predicts Y90-RE therapy response in patients with primary and secondary liver cancers, based on lesion’s invariant texture and shape imaging features extracted from pretherapy FDG PET/CT scans. Our approach utilized computer vision techniques of 3DSG, WKS, and 3DZDs to find regions of tumors that were similar locally even if they have global differences. Using SSBoF to describe imaging features as visual words, we were able to find unique image features that were common within responders and nonresponders. We showed the benefit of using the invariant techniques that were resistant to change in scale, transform, and rotation in assessing Y90-RE treatment response in comparison with routine clinically used image features such as SUV_max and tumor volume. The model improved when combining image features that use both invariant texture and shape features with tumor volume, indicating the value in exploring techniques that combine routine radiological tools with computer-aided higher image resolution techniques. Although our model will need further validation on a large data set, the proposed method is general and can be potentially applied to any lesion from a different disease model.

Abbreviations

AUC

area under the curve

CT

computed tomography

3D

3-dimension

3DGLCM

3D gray level co-occurrence matrix

3DSG

3D spherical Gabor

3DZD

3D Zernike descriptor

EORTC

European Organization for Research and Treatment of Cancer

FDG PET/CT

fluorine-18-2-fluoro-2-deoxy-d-glucose positron emission tomography and computed tomography

MNB

multinomial naive Bayes

PERCIST

PET Response Criteria in Solid Tumors

RECIST 1.1

revised Response Evaluation Criteria in Solid Tumors

SSBoF

spatially sensitive bag of features

SUV

standardized uptake value

SUV_max

maximum standardized uptake value

WKS

wave kernel signature

Y90-RE

yttrium-90 radioembolization.