MRI Radiomics Features of Mesorectal Fat Can Predict Response to Neoadjuvant Chemoradiation Therapy and Tumor Recurrence in Patients with Locally Advanced Rectal Cancer

Abstract

Objective

To interrogate the mesorectal fat using MRI radiomics feature analysis in order to predict clinical outcomes in patients with locally advanced rectal cancer.

Methods

This retrospective study included patients who underwent neoadjuvant chemoradiotherapy for locally advanced rectal cancer from 2009–2015. Three radiologists independently segmented mesorectal fat on baseline T2-weighted axial MRI. Radiomics features were extracted from segmented volumes and calculated using CERR software, with adaptive synthetic sampling being employed to combat large class imbalances. Outcome variables included pathologic complete response (pCR), local recurrence, distant recurrence, clinical T-category (cT), post-treatment T category (ypT), and post-treatment N category (ypN). A maximum of eight most important features were selected for model development using support vector machines and five-fold cross-validation to predict each outcome parameter via elastic net regularization. Diagnostic metrics of the final models were calculated, including sensitivity, specificity, PPV, NPV, accuracy, and AUC.

Results

The study included 236 patients (54 ± 12 years, 135 men). The AUC, sensitivity, specificity, PPV, NPV, and accuracy for each clinical outcome were: for pCR, 0.89, 78.0%, 85.1%, 52.5%, 94.9%, 83.9%; for local recurrence, 0.79, 68.3%, 80.7%, 46.7%, 91.2%, 78.3%; for distant recurrence, 0.87, 80.0%, 88.4%, 58.3%, 95.6%, 87.0%; for cT, 0.80, 85.8%, 56.5%, 89.1%, 49.1%, 80.1%; for ypN, 0.74, 65.0%, 80.1%, 52.7%, 87.0%, 76.3%; and for ypT, 0.86, 81.3%, 84.2%, 96.4%, 46.4%, 81.8%.

Conclusion

Radiomics features of mesorectal fat can predict pathological complete response, local and distant recurrence, as well as post-treatment T and N categories.

Introduction

Magnetic resonance imaging (MRI) is the standard imaging modality for primary staging of rectal cancer and is particularly useful to assess locally advanced rectal cancer (LARC), in which the tumor has spread beyond the muscularis propria and where the locoregional lymph nodes may also be involved. On MRI, the mesorectum, the visceral adipose tissue surrounding the rectum and contained by the mesorectal fascia, is macroscopic spread of tumor into the mesorectum, extramural vascular invasion, presence of mucin, and involvement of mesorectal fascia [1–3]. Several studies in rectal cancer have demonstrated that radiomics, the extraction of vast amounts of quantitative information using computational analysis of imaging, can help predict clinical or prognostic features such as pathologic complete response (pCR) to neoadjuvant chemotherapy, lymph node status, lymphovascular invasion, or synchronous liver metastases [4–12].

An increasing body of evidence points to the critical role of molecular signaling between tumor cells and intratumoral, peritumoral, and even distantly located adipocytes. These molecular interactions can result in elevated levels of inflammatory markers (adipokines) and angiogenic factors such as VEGF or IGF-1, both locally and systemically, promoting tumor growth and metastases [13–15]. Thus, interactions between rectal tumor and mesorectal fat can lead to changes in the molecular profile of adipocytes and subsequently result in subtle changes on MRI not visible to the naked eye, particularly in the early stages of tumor development.

To date, a few studies have performed radiomics assessment on peritumoral tissues on MRI to predict clinical outcomes, treatment response, or prognosis. In patients with breast cancer undergoing dynamic contrast-enhanced MRI, it was reported that combined intra-tumoral and peritumoral radiomics assessment can predict pCR to neoadjuvant chemotherapy [16]. In patients with early stage breast cancer, a positive correlation between peritumoral fat content and axillary lymph nodal status was noted [17]. Recently, radiomics assessment of the rectal primary tumor in combination with the mesorectal compartment on pretreatment MRI was reported to predict patient outcomes with moderate accuracy [18].

We hypothesized that the mesorectal fat contains valuable prognostic information about the tumor and treatment response. Thus, the purpose of this investigation was to interrogate the mesorectal fat using MRI radiomics feature analysis in order to predict clinical outcomes in patients with locally advanced rectal cancer.

Materials and Methods

Study Participants

This retrospective study was approved by the institutional review board with a waiver for informed consent and was compliant with the Health Insurance Portability and Accountability Act. The patient sample was derived from a prior study where a total of 628 patients who were treated at our tertiary cancer care center from June 1, 2009–March 1, 2015 were evaluated to compare different treatment approaches (preoperative chemoradiation followed by postoperative adjuvant chemotherapy versus total neoadjuvant therapy [TNT]) for LARC). From this prior study, we identified only patients who had undergone TNT followed by surgery and who had baseline MRI scans. We excluded patients who had received chemoradiotherapy followed by consolidative chemotherapy as well as patients with no baseline MRI scans, lack of straight T2 axial sequences, studies with too much rectal distension compromising the visibility of mesorectal fat, poor quality or inadequate delineation of mesorectal fascia, or lack of clinical information. A total of 236 patients were included in the final analysis (Figure 1).

Figure 1.

Flowchart of patient inclusion and exclusion. Abbreviations: LARC, locally advanced rectal cancer; MRI, magnetic resonance imaging; TNT, total neoadjuvant chemotherapy.

Treatment

All patients underwent induction-type TNT which typically included folinic acid, fluorouracil and oxaliplatin (mFOLFOX6) for eight cycles; capecitabine and oxaliplatin (CAPOX) for 5 cycles; or weekly fluorouracil/leucovorin and biweekly oxaliplatin (FLOX) [19–22]. Chemoradiotherapy commenced two to three weeks after the completion of induction chemotherapy and included 25–28 radiotherapy fractions with concurrent fluorouracil or oral capecitabine. All patients in this study underwent curative intent total mesorectal excision, and surgical histopathology served as the reference standard for T and N stages.

Imaging Acquisition and Segmentation

The study included baseline MRI scans performed at our institute or outside institutions. Scanner details and MRI parameters are presented in Supplemental Table 1 and Table 1, respectively. The straight axial T2 sequence was used for segmentation. The proximal-most axial image for contouring was defined as where the anterior peritoneal reflection was attached to the rectal wall, identified as a V-shaped structure on the axial plane (seagull sign) [1]. The distal-most image was defined as where the mesorectal fat was still visible above the intersphincteric plane. Three radiologists (VP, VJ, and RB) with 7, 9, and 2 years of experience in oncological imaging, blinded to each other’s contours and to clinical information, individually segmented the mesorectal fat on the T2 axial sequence using Gold LX v2.3.0 software (Hermes Medical Solutions, Inc.). Thirty cases were first segmented by all three radiologists to assess inter-reader agreement. The entire rectum including the tumor and the associated tumor volume extending into the mesorectal fat, discrete tumor deposits in the mesorectal fat, radiologically pathological lymph nodes and extramural vascular invasion, were excluded from the segmented volume.

Table 1.

MRI parameters.

Parameter	In-House Scans Median (Range)	Outside Institutions Scans Median (Range)	All Scans Median (Range)
TE (ms)	103 (18–132)	103 (16–132)	103 (16–132)
Echo Train Length	23 (6–33)	23 (12–25)	23 (6–33)
TR (ms)	4300 (500–9160)	4150 (2508–15000)	4221 (500–15000)
Matrix Size	512×512 (256×256 to 768×768)	512×512 (256×256 to 512×512)	512×512 (256×256 to 768×768)
In-Plane Resolution (mm)	0.47 (0.27–1.48)	0.47 (0.27–1.33)	0.47 (0.27–1.48)
Slice Thickness (mm)	3 (2–6)	4 (3–7)	3 (2–7)

Radiomics Analysis

All segmented volumes were exported to MATLAB (version 9.3.0.713579 (R2017b) (The MathWorks, Inc.) for feature extraction. Images were initially resampled to the median in-plane resolution of 0.47 mm × 0.47 mm. Since radiomics analysis was performed in an aggregated 2D fashion, no resampling in the slice direction was performed. All images were reduced to 32 gray levels prior to radiomic feature calculations. Radiomic features were calculated using CERR software [23] which conforms to the international biomarker standardization initiative [24]. One hundred and one features were calculated in six classes (22 first order, 26 based on gray-level cooccurrence matrices, 16 based on run length matrices, 16 based on size zone matrices, 16 based on neighborhood gray level dependence matrices, and 5 based on neighborhood gray tone difference matrices).

Due to the relatively small sample size, especially for the minority class in various clinical outcomes, predictive models were developed in MATLAB using support vector machines (SVM) and 5-fold cross-validation. Within each fold, elastic net regularization (combining lasso and ridge regression) was used to determine up to eight most important radiomics features for each clinical outcome. Adaptive synthetic sampling of the training data in each fold was employed to equalize class sizes [25] and to prevent the subsequent models from potentially classifying all cases as belonging to the majority class. Clinical outcome parameters from the electronic medical record included pCR, local recurrence, distant recurrence, clinical T-category (cT), clinical nodal category (cN), post-TNT N category (ypN), post-TNT T category (ypT), and post-TNT TNM stage (ypTNM). Radiomics features with low ICC values (< 0.6) were excluded from model development.

Statistical Analysis

Univariate analysis via the Mann–Whitney non-parametric test was performed to assess the number of significant radiomics features for each clinical outcome parameter using SPSS version 25 (IBM Corp.). Radiomics features with a p-value of < 0.05 were considered significant. The results of univariate analysis were not utilized in model development. For all final predictive models (across varying outcomes), diagnostic metrics including sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), accuracy, and area under the curve (AUC) were calculated using MedCalc version 15.0 (MedCalc Software Ltd).

Inter-reader agreement using Jaccard indices (intersection size over union size) and Dice similarity coefficients (DSCs) (twice the intersection size over the sum of two individual regions) were determined on a per-patient basis as well as per-slice basis using MATLAB. Intra-class correlation coefficients (ICCs) were calculated based on per-patient basis Jaccard indices and DSCs in MATLAB using a two-way mixed effects model with single measures. Results from these three metrics were interpreted as follows: values < 0.40 = poor agreement, 0.41–0.59 = fair agreement, 0.60–0.79 = good agreement, and 0.80–1.00 = excellent agreement.

Results

Patient Characteristics

A total of 236 patients (mean age 54 ± 12 years) were included in the final analysis. Patient characteristics including clinical outcomes are presented in Table 2. Supplemental Table 2 details the case distribution with respect to magnetic field strength and indicates that for all clinical outcome measures, cases are evenly distributed across 1.5T and 3T scanners.

Table 2.

Patient characteristics.

Patient Characteristic	Male	Female	Total
Total	135	101	237
Mean age ± SD	54 ± 12	54 ± 13
cT Category
< 4	121	80	201
4	14	21	35
cN
Positive	119	94	213
Negative	16	7	23
cTNM Stage
2	16	7	23
3	119	94	213
pCR
Negative	113	82	195
Positive	22	19	41
Local Recurrence
Yes	25	17	42
No	96	68	164
Distant Recurrence
Yes	22	13	35
No	99	72	171
ypT
0	24	22	46
> 0	111	79	190
ypN
0	101	75	176
> 0	34	26	60
Circumferential Resection Margin
Positive	4	7	11
Negative	131	94	225
Distal Resection Margin
Positive	6	1	7
Negative	129	100	229

cT, clinical T category; cN, clinical N category; cTNM, clinical TNM stage; pCR, pathologic complete response; ypT, post-treatment T category; ypN, post-treatment N category

Inter-Reader Agreement

The average Jaccard index and DSC for inter-reader agreement on a per-patient basis were 0.464 ± 0.128 and 0.616 ± 0.137, respectively (Figure 2). The inter-reader agreement improved when the scores were calculated based on a per-slice basis (Jaccard index and DSCs 0.611 ± 0.111 and 0.750 ± 0.091, respectively).

Figure 2.

Inter-reader agreement scores on two different patients. A, B and C; Segmentation of mesorectal fat by reader 1 (A), reader 2 (B), and reader 3 (C) in a 66-year-old female patient with rectal cancer, with a Jaccard index and dice similarity coefficient of 0.647 and 0.783, respectively. D, E and F; Segmentation of mesorectal fat by reader 1 (D), reader 2 (E), and reader 3 (F) in a 48-year-old female patient with rectal cancer, with a Jaccard index and dice similarity coefficient of 0.298 and 0.451, respectively.

ICCs based on per-patient basis average Jaccard index and DSC scores were > 0.6 for 81% (82/101) of radiomics features across readers. A complete list of ICC data has been included in Supplemental Table 3.

Performance of Predictive Models to Predict Clinical Outcomes

Table 3 presents the elastic net regularization selected features for each fold when considering pCR as the clinical outcome. Table 4 shows the diagnostic metrics for all final predictive models.

Table 3.

Elastic net regularization selected radiomics features for each fold when predicting pathologic complete response.

Feature (ranked)	Fold 1	Fold 2	Fold 3	Fold 4	Fold 5
1	range (FO)	range (FO)	lzhgle (SZM)	lzhgle (SZM)	range (FO)
2	lzhgle (SZM)	lzhgle (SZM)	var (FO)	range (FO)	lzhgle (SZM)
3	maximum (FO)	maximum (FO)	dcnNorm (NGLDM)	clustShade (GLCM)	re (RLM)
4	dcnNorm (NGLDM)	clustShade (GLCM)	range (FO)	dcnNorm (NGLDM)	maximum (FO)
5	invVar (GLCM)	entropy (NGLDM)	clustShade (GLCM)	maximum (FO)	clustShade (GLCM)
6	ze (SZM)	meanAbsDev (FO)	strength (NGTDM)	zp (SZM)	P90 (FO)
7	variance (FO)	std (FO)	coarseness (NGTDM)	entropy (NGLDM)	variance (FO)
8	meanAbsDev (FO)	P90 (FO)	contrast (NGTDM)	variance (FO)	dcnNorm (NGLDM)

FO, first order; GLCM, gray level cooccurrence matrix; RLM, run length matrix; SZM, size zone matrix; NGLDM, neighborhood gray level dependence matrix; NGTDM, neighborhood gray tone difference matrix; lzhgle, large zone high gray level emphasis; dcnNorm, dependence count non-uniformity normalized; clustShade, cluster shade; invVar, inverse variance; ze, zone emphasis; meanAbsDev, mean absolute deviation; zp, zone percentage; P90, 90th percentile; std, standard deviation

Table 4.

Diagnostic metrics for final models to predict clinical outcomes. 95% Confidence intervals are presented within parentheses.

Outcome Measure	AUC	Sensitivity	Specificity	PPV	NPV	Accuracy
pCR	0.89 (0.83–0.96)	78.0 (62.4–89.4)	85.1 (79.3–89.8)	52.5 (43.2–61.6)	94.9 (91.2–97.1)	83.9 (78.6–88.4)
Local Recurrence	0.79 (0.71–0.88)	68.3 (51.9–81.9)	80.7 (73.9–86.4)	46.7 (37.6–56.0)	91.2 (86.7–94.2)	78.3 (72.0–83.7)
Distant Recurrence	0.87 (0.79–0.95)	80.0 (63.1–91.6)	88.4 (82.6–92.8)	58.3 (47.3–68.6)	95.6 (91.8–97.7)	87.0 (81.6–91.2)
Clinical T category	0.80 (0.73–0.88)	85.8 (80.0–90.4)	56.5 (41.1–71.1)	89.1 (74.9–85.4)	49.1 (38.5–59.7)	80.1 (74.4–85.0)
ypN Category	0.74 (0.66–0.81)	65.0 (51.6–76.9)	80.1 (73.4–85.7)	52.7 (44.0–61.3)	87.0 (82.5–90.5)	76.3 (70.3–81.6)
ypT	0.86 (0.79–0.93)	81.3 (75.2–86.5)	84.2 (68.8–94.0)	96.4 (92.8–98.3)	46.4 (38.5–54.4)	81.8 (76.3–86.5)

AUC, area under the curve; PPV, positive predictive value; NPV, negative predictive value; pCR, pathologic complete response; ypN, post-treatment N category; ypT, post-treatment T category

On univariate analysis, 11 radiomics features were significantly different between pathological complete responders and non-complete responders. The final predictive SVM model developed with cross-validation obtained a diagnostic accuracy of 83.9% and an AUC of 0.89 (Figure 3). For predicting local recurrence, univariate analysis revealed 36 radiomics features that were significant. The final predictive model obtained a diagnostic accuracy of 78.3% and an AUC of 0.79 (Figure 4). Meanwhile, 32 radiomics features were significant for predicting distant recurrence on univariate analysis and in this case the final predictive model obtained a diagnostic accuracy of 87.0% and an AUC of 0.87 (Figure 5).

Figure 3.

Area under the curve of the final radiomics model to predict pathologic complete response using eight radiomics features, calculated at 0.89 with a confidence interval of 0.83 to 0.96.

Figure 4.

Area under the curve of the final predictive radiomics model to predict local recurrence using eight radiomic features, calculated at 0.79 with a confidence interval of 0.71 to 0.88.

Figure 5:

Area under the curve of the final predictive model to predict distant recurrence using eight radiomics features, calculated at 0.87 with a confidence interval of 0.79 to 0.95.

To predict clinical T category, 12 radiomics features were significant on univariate analysis when comparing patients with cT4 category to those with lower cT category. The final predictive model with eight radiomics features obtained a diagnostic accuracy of 80.1% and an AUC of 0.80. To predict ypN, only one radiomics feature was significant on univariate analysis. In this case, the final predictive model obtained a diagnostic accuracy of 76.3% and an AUC of 0.74. Eight of the nine significant features were forwarded for model development in the ypT category with a resultant diagnostic accuracy and AUC of 81.8% and 0.86, respectively. As pCR positive group is the same as ypTNM stage 0, no separate model development was carried out for ypTNM stage 0 versus all other ypTNM stages.

Finally, none of the radiomics features were significantly different between node negative and node positive patients or between cTNM stage 2 and cTNM stage 3 patients. Therefore, model development was not pursued for these clinical outcomes. Due to low numbers in the minority class, model development was also not pursued for circumferential resection margin involvement (12 cases) and distal resection margin involvement (7 cases).

Discussion

In our study of radiomics features of the mesorectal fat in 236 patients treated with TNT followed by total mesorectal excision, radiomics models were able to significantly distinguish between patients with complete pathologic response and non-responders (AUC = 0.89); patients with local or distant recurrence and those with no recurrences (AUC = 0.79 and 0.87, respectively); patients with cT4 category and those with less than cT4 category (AUC = 0.80); post-treatment node-negative patients and node-positive patients (AUC = 0.74) and patients with post treatment ypT0 category and those with more than ypT0 category (AUC = 0.86).

Our results suggest that valuable information is contained within the mesorectal fat with respect to rectal cancer prognosis and can be derived from the tumor macro- and micro-environment. Nutrient supply and catabolite drainage to and from the normal rectal wall and rectal tumors must travel through the mesorectal fat by way of vessels and lymphatics. Our results indicate that this environment is rich with predictive information, equivalent to the tumor itself [18]. Rectal cancer arises in close association with white adipose tissue (mesorectal fat). Recent studies in cancers including colorectal cancer have focused on understanding the role of cancer-associated adipocytes in facilitating tumor growth [14; 15; 26; 27]. Adipocytes can be altered or “remodeled” by various factors directly or indirectly, acquire an activated phenotype, and thereby secrete growth factors (e.g. fibroblast growth factor and vascular endothelial growth factor), adipokines (e.g. leptin), and extracellular matrix remodeling factors. These alterations affect cancer growth and metabolism, angiogenesis, invasion, and even therapeutic and radiotherapy resistance [15; 26–32]. In turn, alterations in the molecular expression profile of the peritumoral adipose tissue can also lead to alterations in MRI signal which may not be appreciable on visual assessment.

Over the last few years, radiomics has emerged as an exciting tool to extract hidden data from clinical imaging [9; 33] and predict clinical outcomes such as treatment response, disease-free survival, or overall survival. However, few radiomics studies have investigated the role of peritumoral tissues in predicting disease outcomes. Braman et al [16] evaluated the texture of intra-tumoral and peritumoral regions on pre-treatment breast DCE-MRI to predict pCR in 117 breast cancer patients, demonstrating that radiomics data accurately differentiated between the responders and non-responders as well as identified unique radiomics profiles of response for different biological groups. With respect to rectal cancer specifically, most radiomics studies have focused on the primary tumor [4; 5] or the lymph nodes [6; 7]. A recent study by Shaish et al [18] investigated the radiomics features of the tumor and of the mesorectal compartment in heterogeneously acquired MRIs of 132 patients with LARC to predict pCR, tumor regression grade, and neoadjuvant rectal score; the best overall performance was achieved when using the mesorectal compartment to predict pCR and when using the combined mesorectal compartment and tumor to predict tumor regression grade and neoadjuvant rectal score.

Our study focused primarily on the mesorectal fat and had a larger cohort that underwent a uniform state-of-the-art modern treatment from a single institution. Our pCR rate (17%) was reflective of what can be expected for patients undergoing optimal treatment [34]. While Shaish et al employed a variety of filtering schemes alongside the original image data, resulting in nearly 1600 calculated features, we employed T2-weighted images only with no filtering, leading to 101 calculated features. We achieved an AUC of 0.89 for pCR using only eight features compared to an AUC of 0.80 using 40 features in their study. With 41 cases in the minority class of complete responders and thus a case-to-feature ratio of approximately five, our model does not appear to be overparameterized. The inter-reader agreement scores on a per-patient basis were moderate and improved comparatively when the scores were calculated on a per-slice basis. This demonstrates that the moderate inter-reader agreement scores were mainly due to the differences in the opinion of individual readers regarding where the mesorectal fat started proximally and ended distally, and to a lesser extent were based on variations in the delineation of the margins of the mesorectal fascia on individual slices. Nevertheless, these scores allude to the fact that mesorectal fascia is not always well-defined and there is unavoidable individual variation in segmented volumes of mesorectal fat. To accommodate these differences, we excluded all features with low ICC (< 0.6) for model development. Even accounting for these adjustments, the radiomics models in our study were still able to distinguish between different clinical outcomes with statically significant results. This indicates that the underlying changes in the mesorectal fat are reproducible despite individual differences in the segmented volumes. The ease of segmentation of a relatively large quantity of readily recognizable mesorectal fat, irrespective of somewhat poorly defined mesorectal fascia, makes it a favorable target for future implementation of auto segmentation.

Incremental information from our study concerns not only the value of radiomics to predict pCR but also its value to predict other important clinical outcomes such as local or distant recurrence, cT4 disease, ypN and ypT status. However, we could not demonstrate a significant difference between clinical node positive and node negative disease or clinical TNM stage 2 and stage 3 disease. Due to low numbers, circumferential resection margin and distal resection margin were not evaluated. Interestingly, our model that assessed the radiomics characteristics of mesorectal fat to predict pCR has a very similar classification accuracy to the model by Horvat et al [5] that assessed the radiomics characteristics of rectal tumors. This suggests that the heterogeneity of both the tumor and surrounding mesorectal fat may have similar predictive power.

The clinical implications of our work are several: while the focus of rectal cancer radiomics studies in recent years has been to analyze the primary tumor only, we show that the tumor microenvironment could be the next province of detailed investigation as in other cancer types like breast cancer [17]. Furthermore, the large quantity of fat on MRI and its signal characteristics may allow for easier and more reliable manual and possibly automated segmentation, especially compared with the difficult-to-recognize treated tumor.

There are limitations to our study. Due to the limited sample size, cross-validation was employed rather than splitting the data into training and test cohorts. Clearly, external validation of the predictive models developed is desired and is the next step in this direction. There was class imbalance between various outcome groups, most notably between cases with a positive circumferential resection margin or distal resection margin and their negative equivalents. Class imbalance was addressed by using adaptive synthetic sampling, but this is reliant on the samples available being fully representative of the minority class. As this could not be considered the case for the resection margin data, model development was deemed infeasible. We studied only patients with LARC where there is involvement of mesorectal fat. Whether the same results apply to early stage disease (Stage T1 and T2) without lymph node involvement or tumor invasion beyond the rectal wall has not been assessed but this is another intriguing question of potential clinical importance. It is also important to note that MRI was performed both within our institution and outside, resulting in protocol variations; however, all images were reduced to 32 gray levels and resampled to the same in-plane resolution prior to radiomics feature extraction and all patients received uniform, state-of-the-art TNT treatment. Finally, this was a retrospective study, but with availability of accurate clinical data, we feel that our results can be considered robust.

To our knowledge, this is the largest study to date of the mesorectal fat using MRI radiomics. Our data demonstrates that MRI radiomics-based information derived from mesorectal fat alone can potentially predict various clinical outcomes in rectal cancers, such as pCR and local and distant recurrences, adding to the growing body of evidence for the integral role of mesorectal fat in rectal cancer pathogenesis. Further steps to advance our findings, which we have begun, include validation of our models in an independent cohort, correlation of these models with available gene expression profiles of the mesorectal fat from surgical specimens and testing out deep learning for automated segmentation to facilitate rapid workstation processing and production of a “fat score” type of assessment after completion, to further inform and personalize the treatment of patients with rectal cancer.

Supplementary Material

1776289_Sup_material

Supplemental Table 1. Scanner Details

Supplemental Table 2. Distribution of cases with magnetic field strength

Supplemental Table 3. A complete list of ICC data of radiomics features across readers.

KEY POINTS.

1. Mesorectal fat contains important prognostic information in patients with locally advanced rectal cancer (LARC).

2. Radiomics features of mesorectal fat were significantly different between those who achieved complete vs incomplete pathologic response (accuracy 83.9%, 95% CI: 78.6%–88.4%).

3. Radiomics features of mesorectal fat were significantly different between those who did vs did not develop local or distant recurrence (accuracy 78.3%, 95% CI: 72.0%–83.7% and 87.0%, 95% CI: 81.6%–91.2% respectively).

ABBREVIATIONS

AUC

area under the curve

CAPOX

capecitabine and oxaliplatin

cN

Clinical nodal Category

cT

Clinical T Category

DSCs

Dice similarity coefficients

FLOX

fluorouracil/leucovorin and biweekly oxaliplatin

ICC

Intra-class correlation coefficients

LARC

Locally advanced rectal cancer

MRI

Magnetic resonance imaging

NPV

negative predictive value

pCR

Pathologic complete response

PPV

predictive value

SVM

support vector machines

TNT

Total neoadjuvant treatment

ypN

Post TNT nodal Category

ypT

Post TNT tumor Category

ypTNM

Post TNT TNM Stage