Impact of intratumor heterogeneity on diffusion-weighted MRI as a prognostic indicator in lower rectal cancer

Diffusion-weighed MRI (DWI) has the potential to reveal intra-tumor pathological heterogeneity consisting of the stroma. The present study investigated the value of intra-tumor heterogeneity evaluated by DWI for predicting the survival of patients with lower rectal cancer (LRC). A total of 172 LRC patients underwent radical surgery between 2009 and 2017. Patients with T1 tumors, distant metastasis, and no pre-operative MRI were excluded. Fifty-eight cases were targeted. Intra-tumor heterogeneity on DWI was quantied by an image analysis (MRI heterogeneous score of DWI: mrHSD). All enrolled cases were divided into two groups (high and low groups) according to median mrHSD.


Results
Median mrHSD was 0.457 (0.170-0.823), with a higher score indicating a more heterogeneous pattern on DWI. The frequency of a clinical diagnosis of lymph node metastasis and extramural vascular invasion was higher in the high group. Three-year overall survival (OS) and relapse-free survival (RFS) rates were signi cantly higher in the low group than in the high group (OS: 100% vs 84.6%, p = 0.012; RFS: 85.9% vs 55.6%, p = 0.027). A univariate analysis of RFS showed that the RFS rates of poorly differentiated LRC, a positive circumferential resection margin on MRI, and high mrHSD were worse (p = 0.097, 0.086, and 0.049). A multivariate analysis revealed that high mrHSD tended to be an independent factor for predicting post-operative recurrence (HR: 2.836, p = 0.060).

Conclusion
The quantitative evaluation of intratumor heterogeneity on DWI has potential as an imaging biomarker for predicting post-operative recurrence.

Background
Over the past few decades, multimodal therapies have been performed for advanced lower rectal cancer (LRC) in order to decrease the rates of post-operative recurrence and a positive circumferential resection margin (CRM) [1,2]. The indications for these treatments are generally decided by pre-operative TNM staging [3] based on several conventional modalities, such as endoscopy, endoscopic ultrasound (EUS), computed tomography (CT), and magnetic resonance imaging (MRI). However, diagnostic accuracy has not been satisfactory for providing the appropriate treatment for LRC patients [4,5]. Therefore, the development of a novel diagnostic method that more accurately predicts treatment outcomes and survival than the TNM system diagnosed by conventional modalities is needed.
In recent years, radiomics has been attracting increasing attention in the development of novel diagnostic biomarkers [6]. This research approach in oncology aims to assess tumor heterogeneity quantitatively through an analysis of medical images using histograms, textures, shapes, and mathematical formulae. The clinical signi cance of quantitative values calculated by radiomics, such as the prediction of survival, metastasis, treatment outcomes, and postoperative recurrence, was previously con rmed [7,8]. We also reported that the quantitative value of the intratumor heterogeneity of LRC calculated by an image analysis using a mathematical formula of diffusion images of pre-operative MRI (DWI) was associated with pathological stromal heterogeneity; tumor invasion depth and lymph node metastasis (N(+)) may also be predicted using this value [9]. Radiomics has the potential to predict a number of tumor characteristics based on the tumor structure.
Previous studies identi ed stromal heterogeneity as a prognostic or therapeutic biomarker, and showed that a heteronomous stromal pattern in pancreatic and breast cancers correlated with a poor outcome [10,11]. The tumor stroma ratio is used as a prognostic marker in gastric and colon cancers [12,13]. Therefore, we hypothesized that the quantitative value of intratumor heterogeneity on DWI is also associated with long-term oncological outcomes based on previous studies and our ndings. The present study investigated the value of tumor heterogeneity on DWI images for assessing long-term oncological outcomes in patients with LRC.

Study design and patients
This was a respective study that investigated the clinical value of intra-tumor heterogeneity evaluated by DWI in LRC. The present study included patients with rectal adenocarcinomas that were mainly located under the peritoneal re ection who underwent curative surgery with regional lymph node dissection between April 2009 and March 2017 at the Division of Digestive Surgery of Kyoto Prefectural University of Medicine (KPUM). Exclusion criteria were as follows: (1) patients diagnosed with T1 LRC before treatment initiation; (2) patients diagnosed with distant metastases before treatment initiation; (3) patients who did not undergo MRI with DWI before treatment and surgery. One hundred and forty-six cases ful lled the inclusion criteria and 114 were subsequently excluded; therefore, data collected from 58 LRC cases were analyzed in the present study ( Figure 1).

Surgical procedures
All surgeries were performed or supervised by surgeons with su cient experience and certi cated by The Japan Society of Endoscopic Surgery and The Japanese Society of Gastrointestinal Surgery. The surgical procedure was selected based on the Japanese colorectal cancer guidelines [14]. Lateral lymph node dissection (LLND) was performed for LRC with 5-mm or larger lateral lymph nodes in the short axis diameter in pre-treatment CT or MRI. Pre-operative CRT was performed for LRC clinically diagnosed as T3 or N(+) in order to secure CRM.

Clinical and pathological diagnoses
Clinical and pathological TNM staging was performed according to the Japanese Classi cation System [3]. The clinical diagnosis of LRC was based on the ndings of the conventional modalities, endoscopy, contrast-enhanced CT, and MRI. The pathological diagnosis was decided by at least two experienced pathologists and evaluated according to the Japanese Classi cation System. Brie y, lymph nodes in the mesorectum, around the inferior mesenteric artery, and in both lateral areas were de ned as the regional lymph nodes of LRC. The tumor differentiation grade was classi ed according to the differentiated type, which was the main cellular component of tumors.
Pre-operative and pre-treatment MRI The MRI protocol employed was similar to that in our previous study [9]. Imaging was performed with a 1.5 or 3.0 T pelvic MRI with pelvic phased-array coils at KPUM or related medical centers. T2-weighted axial images with a section thickness of 5-7 mm and sagittal images with fast spin-echo sequences were acquired. An axial diffusion-weighted sequence with background body signal suppression (DWIBS, b-values 800-1000 s/mm 2 ) was also obtained. Cases without pre-operative CRT underwent MRI for staging within 2 months before surgery, and those with pre-operative CRT underwent MRI for pretreatment staging and re-staging in order to diagnose therapeutic responses almost 4 to 7 weeks after the completion of CRT. Positive CRM evaluated by MRI (mrCRM) was de ned as a tumor, lymph node, extramural vascular invasion (EMVI), or tumor deposit located <1 mm from the mesorectal fascia [15,16]. EMVI evaluated by MRI (mrEMVI) was diagnosed according to the 5-scale EMVI scoring system suggested by Smith et al. [17], and scores of 3 and 4 were de ned as positive mrEMVI. MRI ndings were interpreted by a digestive surgeon based on the radiological reports of experienced radiologists.

Quanti cation of intratumor heterogeneity on DWI images
The intra-tumor heterogeneity of DWI images was quanti ed as described in our previous study [9]. The heterogeneity was evaluated in only MRI before treatment initiation. Brie y, DWI were evaluated for the maximum cut surface of an axial image of the rectal tumor, which was identi ed by a T2-weighted axial image ( Figure 2A). The distribution of signal intensities in this high-intensity area was evaluated, and the maximum (MAX) and minimum (MIN) values of signal intensities in the tumor were measured ( Figure 2B, C). The intra-tumor heterogeneity of the signal intensity on DWI was quanti ed using the following

Statistical Analysis
Comparisons were performed between both groups using the Mann-Whitney U test or Fisher's exact test. P values <0.05 were considered to be signi cant. Survival curves were constructed using the Kaplan-Meier method, and differences in survival were examined using the Log-rank test. Uni-and multivariate analyses of the factors in uencing survival were performed using Cox's proportional hazard model. Differences were considered to be signi cant when the relevant p value was <0.1 in Cox's proportional hazard model. These analyses were performed using JMP statistical software (JMP version 10).

Patient characteristics and MRI ndings
The clinicopathological characteristics and MRI ndings of 58 cases were summarized in Table 1 Clinical factors related to intratumor heterogeneity on DWI in LRC All enrolled cases were divided into two groups according to median mrHSD (mrHSD: 0.457): high and low groups. The relationships between mrHSD and clinical factors were statistically analyzed ( Table 2). The number of cases diagnosed with regional N(+), lateral N(+), and positive mrEMVI were signi cantly higher in the high group than in the low group. Furthermore, a higher number of cases in the high group underwent preoperative CRT than in the low group. This result was attributed to the high group containing more clinical N(+) and positive mrEMVI cases.
The value of intratumor heterogeneity on DWI for predicting patient survival Kaplan-Meier overall survival (OS) and relapse-free survival (RFS) curves were calculated, and the mrHSD status was compared ( Figure 3). All LRC cases in the low group survived three years after surgery, and OS rates were signi cantly higher in the low group than in the high group (3-year OS: 100% vs 84.6%, p=0.012) ( Figure 3A). Moreover, RFS rates were signi cantly higher in the low group than in the high group (3-year RFS: 85.9% vs 55.6%, p=0.027).
We then performed uni-and multivariate analyses of 3-year RFS. The univariate analysis showed that the 3-year RFS rates of cases with non-differentiated LRC, LRC cases diagnosed with positive mrCRM, and the mrHSD high group were worse (p=0.097, 0.086, and 0.049). The multivariate analysis of signi cant factors (p<0.1) revealed that mrHSD tended to be an independent factor for predicting post-operative recurrence (HR: 2.836, 95%CI: 0.939-10.314, p=0.060).

Discussion
The diversity of stromal distribution in individual cancer tissues was previously demonstrated in the pathological ndings of digestive malignant tumors [18,19]. Highly differentiated cancer cells generally proliferate well, resulting in extensive areas of the tumor being occupied almost exclusively by these cells. In contrast, poorly differentiated cancer cells grow invasively while forming a stromal region around cancer nests. Hence, poorly differentiated tumor tissues mostly contain stromal regions. Furthermore, the mixture of cancer nests with these different histologies has frequently been reported, and the pattern of stromal distribution in histologically mixed tumors is generally heterogeneous. Previous studies reported that the heterogeneity of stromal distribution in pathological ndings has potential as a prognostic marker [10,11], suggesting that the outcomes of patients may be estimated in the pre-operative setting by evaluating the cancer stromal pattern indirectly using imaging modalities.
DWI images of MRI quantitatively reveal the diffusional restriction of water molecules in the stroma of tissues. Previous studies reported that the distribution of stroma in tissue may be predicted using DWI images [20,21]. In malignant tumors, the stromal region of cancer tissue is generally smaller than that of normal tissue because of the proliferation of cancer cells and their invasion of normal tissue, suggesting that the diffusional movement of water molecules is restricted in cancer tissues. Therefore, solid malignant tumors were observed as a space-occupying region with high intensity in DWI images of MRI. The e cacy of DWI images of MRI for diagnosing primary tumors, lymph node metastasis, and distant metastasis, is similar to that of positron emission tomography-computed tomography (PET-CT) [22].
Due to the superiority of DWI for evaluating the stroma, the complexes and heterogeneities of stromal structures may be predicted using a DWI image analysis. The heterogeneous pattern of stroma in pathological ndings was almost consistent with the uniformity of the pathological stromal pattern in our previous study [9]. Based on these ndings, we investigated whether the intratumor heterogeneity of DWI was associated with patient survival.
The present results demonstrated that the intratumor heterogeneity of DWI was not associated with tumor histology; however, a correlation with the cN stage and EMVI was observed ( Table 2). Although the number of cases that underwent pre-operative CRT was signi cantly higher among those with highly heterogeneous tumors, the RFS rate of these cases was signi cantly lower than that of the low heterogeneity group. Furthermore, histology and CRM before treatment initiation were associated with RFS. The multivariate analysis identi ed tumors with a highly heterogeneous pattern as an independent risk factor for predicting poor RFS. These results suggest that a novel treatment strategy needs to be considered for patients with highly heterogeneous tumors on DWI images, and also that intratumor heterogeneity on DWI may be used as an imaging biomarker for predicting recurrence and planning individual treatment strategies. In the present study, we evaluated intra-tumor heterogeneity after CRT; however, a relationship with clinical factors or patient outcomes was not observed (data not shown). This image analysis has recently been performed in clinical settings using arti cial intelligence technology (AI); therefore, the development of detailed evaluations of intratumor heterogeneity by AI is expected in the future [23].
There were some limitations that need to be addressed. The present study was retrospective with a small sample size, which may have limited the statistical power. DWI are often affected by the background on T2-weighted data, which is known as "T2 shine-through" [24]. To avoid this interference, an apparent diffusion coe cient map (ADC) was used. However, ADC data were not obtained because ADC was not routinely made. Inter-observer variability in the interpretation of clinical images also needs to be evaluated; however, the image analysis in the present study was conducted by one clinician. Although the present study had these limitations, the results obtained are of clinical signi cance.

Conclusions
The quantitative evaluation of intratumor heterogeneity on DWI images has potential as an imaging biomarker for predicting post-operative recurrence. Although the present results need to be validated in studies with large sample sizes, they may contribute to improvements in the pre-operative evaluation of LRC.