Radiomic-based prediction of microsatellite instability in colorectal cancer at initial computed tomography evaluation

Abstract

Purpose:

To predict microsatellite instability (MSI) status of colon cancer on preoperative CT imaging using radiomic analysis.

Methods:

This retrospective study involved radiomic analysis of preoperative CT imaging of patients who underwent resection of stage II–III colon cancer from 2004–2012. A radiologist blinded to MSI status manually segmented the tumor region on CT images. 254 intensity-based radiomic features were extracted from the tumor region. Three prediction models were developed with: 1) only clinical features, 2) only radiomic features, and 3) “combined” clinical and radiomic features. Patients were randomly separated into training (n=139) and test (n=59) sets. The model was constructed from training data only; the test set was reserved for validation only. Model performance was evaluated using AUC, sensitivity, specificity, PPV, and NPV.

Results:

Of the total 198 patients, 134 (68%) patients had microstatelite stable (MSS) tumors and 64 (32%) patients had MSI tumors. The combined model performed slightly better than the other models, predicting MSI with an AUC of 0.80 for the training set and 0.79 for the test set (specificity=96.8% and 92.5%, respectively), whereas the model with only clinical features achieved an AUC of 0.74 and the model with only radiomic features achieved an AUC of 0.76. The model with clinical features alone had the lowest specificity (70%) compared with the model with radiomic features alone (95%) and the combined model (92.5%).

Conclusions:

Preoperative prediction of MSI status via radiomic analysis of preoperative CT adds specificity to clinical assessment and could contribute to personalized treatment selection.

Introduction

Colon cancer is the third most common cancer and the second most common cause of cancer mortality in men and women combined worldwide; however, not all colon cancers are genetically similar. Recently, colon cancers with microsatellite instability (MSI – about 15%) and the remaining microsatellite stable (MSS) tumors have been highlighted in clinical practice to have unique clinical and pathologic characteristics and distinct responses to treatment. MSI colon cancers are caused by loss of DNA mismatch repair activity and have a more indolent course and a slightly better prognosis when compared to MSS cancers [1]. MSI colon tumors are typically observed in younger patients, frequently arise in the right colon, have increased lymphocytic infiltrate, are associated with Lynch Syndrome (3% of all colon cancers and 3% of MSI tumors), and respond favorably to immune check point blockade [2]. MSS cancers on the other hand are more likely to occur in older patients and in non-hereditary conditions, arise in the left colon [3], and are treated with standard-of-care 5-FU based chemotherapy [4].

MSI status is currently assessed preoperatively by colonoscopic biopsy requiring a team of anesthesiologists, gastroenterologists, and various nursing and support staff for the procedure which can be costly. This procedure is the standard of care but has two challenges. First, biopsy is an invasive test in which the DNA extracted from the sample may not meet the minimum quality/quantity criteria for the genetic assay, thus resulting in unknown MSI status [5]. Second, many patients with colon cancer are first diagnosed via CT imaging rather than colonoscopy when they become symptomatic (such as with pain or bloating) and seek medical attention. In particular, this may be seen in younger patients who are more likely to have MSI tumors, as they are too young to meet the current United States Preventive Services Task Force screening recommendation for colonoscopy, which is for patients over 50 years of age [6], or the new American Cancer Society guidelines of 45 years of age [7]. These patients then experience a treatment delay while waiting for either a colonoscopy or surgery to determine MSI status as these tumors are currently indistinguishable by a radiologist on routine CT imaging.

The approach to the treatment of many luminal gastrointestinal cancers (esophageal [8; 9], gastric [10], rectal [11]) is migrating towards a neoadjuvant paradigm. Correspondingly, preoperative chemotherapy for locally advanced operable primary colon cancer has been found to be feasible, safe, and capable of inducing significant downstaging according to preliminary results from the FOxTROT Collaborative group [12]. In this context, the early detection of MSI status could enable risk reduction in patients with MSI tumors who would benefit most from immunotherapy; at the same time, side effects from 5-FU based chemotherapy, which has limited value in MSI colon cancers, could be avoided, and treatment could be better individualized. One approach to enabling early detection is to leverage computational analyses of high resolution routine cross-sectional imaging using radiomics.

Radiomics is a burgeoning field and has shown promising results in distinguishing a complete response from residual tumor after total neoadjuvant treatment in rectal cancer [13], estimating disease free survival in early stage lung cancer [14], assessing tumor recurrence risk in breast cancer [15], and assessing Gleason score in prostate cancer [16]. Given the ubiquity of CT in the preoperative evaluation and TNM staging of colon cancer and the potential gap in knowledge of a tumor’s MSI status, we hypothesized that radiomic features extracted from portal venous phase CT could identify MSI colon cancers. Therefore, the purpose of this study was to design a prediction model based on radiomic features extracted from preoperative CT to non-invasively identify MSI colon cancers.

Methods

Patient Selection

In this Health Insurance Portability and Accountability Act-approved retrospective study, a waiver of informed patient consent was obtained from our institutional review board. A surgical database was searched for consecutive patients that underwent resection of stage II–III colon cancer between January 2004 and December 2012. Patients with 90-day postoperative deaths and another stage IV malignancy were excluded. Thus, the search yielded 981 patients. Patients with known mismatch repair proteins expression status determined using immunohistochemistry (IHC) were considered eligible for our study (n = 328). Among these patients, those with a preoperative portal-venous phase CT scan of the abdomen and pelvis within 8 weeks of surgery available on the institutional picture archiving and communication system (Centricity PACS, GE Healthcare) were included (n = 219). Finally, patients with CT images of poor quality or radiographically occult tumor were excluded (n = 21). Therefore, the final cohort consisted of 198 patients with stage II or III colon cancer with IHC-proven MSI status. During the study period, IHC testing was routinely performed on specimens from every patient ≤ 50 years old undergoing a colectomy for colon cancer at our institution and was selectively performed in patients > 50 years old who had a family history of colorectal cancer and/or clinical features suspicious for MSI. Data regarding demographics and clinicopathologic features were collected by review of the electronic medical record.

Clinical and Pathological Variables

Within the Department of Pathology, after standard gross and histopathologic assessment was performed, IHC for mismatch repair proteins expression was performed using the standard streptavidin-biotin-peroxidase procedure. Primary MoAbs against MLH1 (clone G168–728, diluted 1:250 (PharMingen®)), MSH2 (clone FE11, diluted 1:50 (Oncogene Research Products®)), MSH6 (clone GRBP.P1/2.D4, diluted 1:200 (Serotec Inc®)), and PMS2 (clone A16–4, diluted 1:200 (BD PharMingen®)) were used. Tumors deficient in MLH1, MSH2, MSH6, or PMS2 proteins were used as external controls. Tumors that showed a total absence of nuclear staining while their adjacent benign tissue showed nuclear staining were scored negatively for the expression of that protein and were deemed MSI [17].

CT Acquisition

Patients included in the study were scanned as part of routine clinical care. As a result, scans were acquired at our center as well as outside institutions with images uploaded to PACS. CT acquisition was performed using a variety of CT manufacturers including: Elscint (2), GE (121), Philips (16), Siemens (44), Toshiba (14), and unknown (1). Mean acquisition parameters were: 120 kVp (range: 100–140 kVp), exposure time 751 ms (range: 500–1782 ms), and tube current 333 mA (range: 100–752 mA). Images were reconstructed at a slice thickness varying from 1–7.5 mm (mean: 4.6 mm) with a reconstruction diameter range from 274–500 (mean: 400). All contrast agents (for example: Omnipaque, Isovue, Optiray, and Ultravist), contrast rates, and dosages were acceptable.

Radiomic Analysis

CT scans were transferred from PACS to a local workstation for advanced image analysis. A radiologist with expertise in body imaging manually segmented the tumor region on all CT slices using ITK-SNAP software; segmentation was then secondarily reviewed and edited by a second body imaging radiologist. Both radiologists were blinded to MSI status. A total of 254 well-established intensity-based radiomic features were extracted from the tumor region using grey-level co-occurrence matrices (GLCMs) [18], run-length matrices (RLMs) [19], local binary patterns (LBPs) [20–22], fractal dimension (FD) [23; 24], intensity histogram (IH) and angle co-occurrence matrices (ACMs) [25; 26]. These features included 19 features from GLCM, ACM1, and ACM2, respectively; 11 from RLM, 5 from IH; 127 from LBP; and 54 from FD [27; 28]. Each feature was extracted from each CT slice and averaged to obtain a single value for each patient. These features quantify heterogeneity in CT enhancement patterns. GLCM features, for example, encode the spatial distribution of two neighboring pixels located at a distance and direction, whereas RLM features represent the coarseness of an image via counting the number of consecutive pixels. FD features capture the self-similarity present within the image. ACMs describe the orientation patterns present in the tumor region. Software created in MATLAB R2015a (The MathWorks, Inc.) was used to extract the radiomic features. These radiomic features have been previously described in the literature [25].

Statistical Analysis of Clinicopathologic Variables

The significance of associations between MSI and clinicopathologic variables was assessed using the Fisher’s exact or Chi square test for categorical variables, and the Student’s t-test or Mann-Whitney test, as appropriate by the type of distribution, for continuous variables. A p-value ≤ 0.05 was considered significant. Statistical analyses of clinicopathologic variables were performed using statistical software (Statistical Package for the Social Sciences, Inc., SPSS software version 25).

Prediction Model Building

Three MSI prediction models were designed using our available preoperative data: (1) prediction model using clinical variables, (2) prediction model using radiomic features extracted from the tumor region in CT, and (3) prediction model using a combination of clinical variables and radiomic features. The first prediction model included clinical variables significantly associated with MSI on univariate statistical analysis. The second prediction model included radiomic features that were significant on the Wilcoxon rank-sum test. In this prediction model, highly correlated radiomic features (i.e., features with correlation coefficient > 0.9) were removed from the feature set to reduce the number of variables. Finally, in the third prediction model, clinical variables were combined with radiomic features. A random forest classifier was utilized to create all three MSI prediction models. Random forest is considered as one of the most accurate general-purpose machine learning techniques that handles a large number of features without overfitting [29]. The patient cohort was randomly divided into training and test sets to ensure that the prediction models were evaluated on independent data: 70% of the data was used to design the prediction model, while the remaining 30% was used to evaluate the performance of the model. The performance of the models was evaluated using the following: AUC under the receiver operating characteristic curve, sensitivity, specificity, PPV, and NPV. The random forrest models were designed in MATLAB.

Results

Patients

A total of 198 patients met the inclusion criteria and were included in the analysis. Figure 1 shows the patient selection. CT scans performed at outside institutions were reviewed by a radiologist for adequate contrast enhancement and overall image quality, and 21 patients were excluded due to poor image quality. Reasons for this exclusion included: no or poorly timed (arterial instead of portal venous phase) contrast (n = 7), slice spacing issues resulting in an inability to download series for transfer to segmentation software (n = 7), unclear or occult tumor (n = 2), breathing artifact resulting in missing or partially duplicated tumor (n = 2), and 1 patient had indistinct borders due to the tumor acting as a lead point for intussusception.

Fig. 1.

Flowchart of the study population. MSS: microsatellite stability, MSI: microsatellite instability

Among the 198 patients, 152 had CT scans performed at outside institutions and 46 had scans performed at our institution; 134 (68%) patients had MSS tumors and 64 (32%) patients had MSI tumors. Table 1 lists clinicopathologic variables stratified by MSI status. MSI status was significantly associated with age, primary tumor location (left or right side), node positivity, number of positive nodes, stage, differentiation, mucinous adenocarcinoma, tumor budding, tumor infiltrating lymphocytes, vascular invasion, and KRAS mutant status on univariate analysis.

Table 1:

Clinicopathological features of patients who underwent surgery for stage II–III colon cancer stratified by MSI status^a

Characteristics	MSS (n = 134)	MSI (n = 64)	p value
Median age, years (range)	51.5 (27–85)	62.0 (29–86)	0.007
Gender			0.84
Female	67 (50.0)	33 (51.6)
Male	67 (50.0)	31 (48.4)
Primary tumor location			< 0.05
Right	63 (47.0)	58 (90.6)
Left	71 (53.0)	6 (9.4)
T stage			0.55
T1	3 (2.2)	1 (1.6)
T2	10 (7.5)	2 (3.1)
T3	100 (74.6)	53 (82.8)
T4	21 (15.7)	8 (12.5)
Node-positive primary
No	62 (46.3)	46 (71.9)	<0.05
Yes	72 (53.7)	18 (28.1)	<0.05
N1	54 (40.3)	12 (18.8)
N2	18 (13.4)	6 (9.4)
Median number of positive nodes (range)	1.0 (0–10)	0.0 (0–8)	< 0.05
Median total number of nodes (range)	23.5 (11–119)	28.0 (6–117)	0.18
CRC stage			< 0.05
II	62 (46.3)	46 (71.9)
III	72 (53.7)	18 (28.1)
Tumor differentiation			< 0.05
Poor	26 (19.4)	29 (45.3)
Moderate	108 (80.6)	35 (54.7)
Well	0 (0.0)	0 (0.0)
Mucinous adenocarcinoma			< 0.05
Yes	37 (27.6)	32 (50.0)
No	97 (72.4)	32 (50.0)
Tumor budding			< 0.05
Yes	40 (29.9)	7 (10.9)
No	29 (21.6)	15 (23.4)
Unknown	65 (48.5)	42 (65.6)
Tumor infiltrating lymphocytes			< 0.05
Yes	16 (11.9)	22 (34.4)
No	60 (44.8)	11 (17.2)
Unknown	58 (43.3)	31 (48.4)
Perineural invasion			0.098
Yes	45 (33.6)	14 (21.9)
No	86 (64.2)	50 (78.1)
Unknown	3 (2.2)	0 (0.0)
Vascular invasion			< 0.05
Yes	78 (58.2)	24 (37.5)
No	56 (41.8)	40 (62.5)
Positive margin			0.15
Yes	0 (0.0)	1 (1.6)
No	134 (100.0)	63 (98.4)
KRAS mutant			< 0.05
Yes	6 (4.5)	4 (6.3)
No	11 (8.2)	13 (20.3)
Unknown	117 (87.3)	47 (73.4)
BRAF mutant			0.17
Yes	5 (3.7)	5 (7.8)
No	14 (10.4)	11 (17.2)
Unknown	115 (85.8)	48 (75.0)
HNPCCb			-
Yes	-	15 (23.4)
Sporadic	-	16 (25.0)
Unknown	-	33 (51.6)

^aValues in the table are numbers of patients (percentages) unless otherwise indicated

^bConfirmed after V600E BRAF mutation and/or MLH1 hypermethylation promoter assays MSI, Microsatellite instability; MSS, Microsatellite stability; CRC, Colorectal cancer; HNPCC, Hereditary nonpolyposis colorectal cancer

Multivariate preoperative MSI prediction models

Of the 198 patients, 139 were included in the training set and 59 were reserved strictly for tesing the performance of the prediction models. The first model included two clinical variables significant on univariate analysis and commonly cited in the literature (age and location of tumor [right or left side]) and achieved an AUC of 0.74 with the test data.

The second model included radiomic features. In total, 93 of 254 (37%) radiomic features showed a significant association with MSI. After removing correlated features, 40 radiomic features were selected for inclusion in the second prediction model, which achieved an AUC of 0.76 with the test data.

The third model combined both clinical variables and radiomic features into a single prediction model. This model achieved an AUC of 0.79 with the test data. The receiver operator curves for both training and test data are demonstrated in Figure 2.

Fig. 2.

Model performance comparison for the training and test data

Model performance was compared between the models using sensitivity, specificity, NPV, and PPV with the training and test data (Table 2). The operating point for these performance metrics was determined using Youden’s index. Radiomic features from all feature categories showed significant performance, with GLCM showing the largest percentage and demonstrating increased homogeneity (Figure 3), followed by RLM and FD (Figure 4).

Table 2:

MSI prediction model results for training and test data

	Training			Test
	Clinical	Radiomic	Clinical + radiomic	Clinical	Radiomic	Clinical + radiomic
AUC	0.75	0.74	0.8	0.74	0.76	0.79
Sensitivity	68.9	28.9	44.4	63.2	31.6	31.6
Specificity	74.5	95.8	96.8	70	95	92.5
PPV	56.4	76.5	87	50	75	66.7
NPV	83.3	73.8	78.5	80	74.5	74

Fig. 3.

Examples of MSI positive and negative tumors classified correctly by our prediction model. (a) shows the cross-sectional axial image of an MSI tumor and (b) shows the tumor at the pixel level. (c) shows the cross-sectional axial image of an MSS tumor and (d) shows the tumor at the pixel level. Homogeneity is measured in MSI tumors which tend to be more indolent than MSS tumors. This is consistent with radiomics of other cancers indicating tumor heterogeneity is a marker for disease aggressiveness

Fig. 4.

Percentage of significant radiomic features in the training data by radiomic feature category

Discussion

We demonstrate that CT radiomic analysis can identify MSI colon cancers on initial pretreatment portal venous phase CT. We achieved similar results with the test and training sets using all three prediction models. Our combination model of radiomic features and clinical features had the best discriminatory ability in both the training cohort (AUC 0.80) and test cohort (AUC 0.79), higher than that of the model with radiomic features alone and the model with clinical features alone (test cohort: AUC of 0.76 and 0.74, respectively). The model with clinical features alone had the lowest specificity (70%) compared with the model with radiomic features alone (95%) and the combined model (92.5%). While our sensitivity for the combined model was low (test 31.6%) compared to the clinical features alone (test 63.2%), it is most likely due to the low sensitivity of the radiomic features (test 31.6%) which is lowering our sensitivity in the combined cohort. At the pixel level, increased homogeneity in the MSI tumors was observed, consistent with radiomics investigations of other cancers where increased tumor heterogeneity on imaging is a biomarker for disease aggressiveness [30].

Our results are of clinical value to medical and surgical oncologists. Since MSI and MSS tumors have different treatment strategies, the ability to predict MSI on preoperative CT scans routinely performed for detection and initial TNM staging of colon cancer may be useful to noninvasively stratify neoadjuvant treatment for colon cancer patients. Currently, MSI and MSS colon cancers are indistinguishable to radiologists on routine CT imaging. However, as standard of care moves from an adjuvant (postoperative) chemotherapy paradigm to a neoadjuvant (preoperative) paradigm for the treatment of colon cancer in higher risk tumors, the radiomics-based prediction model could be clinically impactful to differentiate MSI tumors early on. A model such as ours could serve as an imaging biomarker for MSI and could individualize initial treatment and reduce the risk to patients from unnecessary side effects due to the traditional 5-FU based chemotherapy in the neoadjuvant setting for MSI colon cancers. While not an adequate screening test due to the low sensitivity, our high specificity suggests patients who test positive for MSI could potentially go straight to neoadjuvant immunotherapy, but those who do not will still need to rely on other methods of MSI detection. Additionally, a validated imaging biomarker could decrease the time interval between diagnosis and treatment, and in the future, potentially act as a virtual biopsy.

An important consideration for all patients undergoing tissue biopsy, not just for colon cancer, is the rate of sampling error due to intratumoral heterogeneity or insufficient sample size [31; 32] as well as challenges to practitioners because MSI testing may not be routinely available—all of which may lead to an unknown MSI status. The availability of radiomics assessment of MSI status could aid in diagnosis, prognosis, and selection of optimal treatment for patients on initial CT evaluation. Our results demonstrate the potential of quantitative image analysis to detect MSI with high specificity in colon cancers on routine portal venous CT scan.

While a growing body of radiomics investigations in a wide range of cancers are being published, to our knowledge, our study is the first of its kind on CT radiomics for the prediction of MSI colon cancers. Our results indicate and further support the potential contribution of radiomics to aid in the development of precision medicine for patients with cancer.

This study has several limitations. First, to maximize our sample size, we included CT scans performed at outside institutions which were acquired under varying protocols. Given the potential differences in reproducibility in radiomic features with varying scanners and imaging protocols [33], further studies are needed to see if radiomic features for colon cancers are reproducible under different contrast-enhanced CT conditions. Nevertheless, the predictive potential of our radiomic features despite variation in CT imaging parameters is encouraging because colon cancer, once detected, typically does not undergo further advanced imaging such as magnetic resonance imaging (MRI) because portal venous phase CT remains the standard of care for TNM staging [34]. A sentinel paper on the field of radiomics discussed the intent of conducting radiomics analysis with standard of care images [35]. Our results using varying CT protocols, therefore, may be broadly applicable due to the ubiquity of CT imaging. Another limitation is that as a single-center retrospective analysis, our results need further validation in a prospective multicenter study to evaluate the reproducibility of our prediction models [36]. A third limitation is that manual segmentation may be a source of observer variability. In the future, automated segmentation may be possible which would help remove some of this variability.

Finally, two interesting elements of our patient data set were (1) the high number of MSI tumors at our tertiary care cancer center and (2) the average age of MSI patients was actually older than the MSS counterpart (62 years in MSI cohort compared to 51.5 years in MSS). Review of the literature cites an incidence of MSI in the population of about 15%. In our study, 32% of patients had MSI, which may be partly due to the referral pattern of our large tertiary referral center. While this enriched population may have influenced our results, our patients were accrued consecutively and thus are not necessarily a source of bias. Regarding the older age of MSI patients (62 years) which is contrary to the typically younger MSI CRC population frequently cited in the literature, it is only the patients with MSI CRC due to Lynch syndrome who tend to be younger compared to those with sporadic MSI CRC who tend to be older due to loss of MLH1 expression which increases with age [3]. Our results, therefore, might reflect the balance in the two different MSI populations as sporadic MSI CRC is more common (12%) than Lynch-associated MSI CRC (3%).

In conclusion, our results using a radiomics prediction model to identify MSI colon cancers has important clinical implications for the future of precision medicine. In this work, we show that a radiomics prediction model could enable a diagnosis to be established that influences appropriate preoperative treatment (traditional 5-FU based chemotherapy versus immunotherapy) and this could be highly useful in the coming era of neoadjuvant chemotherapy for colon cancer.