Introduction

Rectal cancer represents a major cause of morbidity and mortality worldwide, accounting for about 44,000 new cases and 51,000 deaths per year in the USA [1]. Most cases are locally advanced, i.e., clinically staged T3–4 or node positive, at the time of diagnosis.

The gold standard for the treatment of locally advanced rectal cancer (LARC), according to the current guidelines, is neoadjuvant chemo-radiotherapy (nCRT) followed by total mesorectal excision (TME). Neoadjuvant therapy has shown to decrease the rate of local recurrences [2], and can lead to complete pathological response (pCR) in some cases [3, 4]. In patients achieving a clinical complete response (cCR) a rectal-sparing approach could be attempted in order to spare surgery, that in case of complete response would represent an overtreatment burdened by the risk of major complications and severe impairment on bowel function and quality of life [3, 5]. Given these premises, it is crucial to identify reliable predictors of tumor response in order to offer a patient-tailored treatment. Several clinical and molecular biomarkers have been investigated as possible predictors of response to treatment [6,7,8,9]. As for imaging, MRI may represent a valuable tool to predict response to nCRT. This technique offers relevant advantages, due to its high resolution in soft tissue analysis and the possibility of a multiparametric approach [10, 11]. MRI T2-weighted imaging, in particular, is used to assess the degree of tumor regression on the basis of morphological changes. To do so, a classification similar to Mandard’s tumor regression grade (TRG) system [12], mrTRG, has been proposed, based on the hypointensity in the T2-weighted sequences of the fibrotic tissue in the lesion [13,14,15]. This classification showed to predict recurrence-free survival (RFS) disease-free (DFS) and overall survival (OS) [16], while in other studies, mrTRG showed a low correlation with pathological TRG [17]. In particular, the inter-observer agreement for this type of evaluation was not satisfying in TRG1 and 2 cases, which are the most relevant in the evaluation of a complete response [17], thus an automated quantitative approach for the evaluation of the T2-weighted signal intensity inside the lesion has been proposed showing a good inter-observer agreement and an excellent correspondence to the histopathological findings [18]. Taylor et al. [19] reported that more than 70% of tumor volume changes detected by T2-weighted images define tumor responses as determined by morphological evaluation by MRI.

However, these findings may underestimate the difficulty to determine the margin between residual tumor and surrounding normal tissues and to differentiate between residual tumor and fibrosis after nCRT, on the basis of morphological changes only. More sensitive and reliable MRI markers are needed to evaluate the efficacy or predict early responses to treatment in clinical practice. Some studies applied texture analysis (TA) in order to predict pCR; different T2-weighted images, DWI and PET/TC features showed a significant correlation with pathological findings [20,21,22,23,24,25] and some T2-weighted images features were found to associate with DFS, RFS, OS [22].

The aim of the present study was to evaluate the correlation of first- and second-order TA features of the primary lesion’s volume in pre- and post-nCRT T2-weighted MRI images and histopathological TRG.

Methods

Patient selection

We retrospectively included consecutive patients with confirmed diagnosis of LARC who underwent nCRT followed by curative surgery at our center between 2013 and 2019. Patients with stage II and II rectal cancer, located at less than 11 cm from the anal verge, who received preoperative radiotherapy with a total dose of 50.4 Gy (in daily fractions of 1.8 Gy five times a week), with concomitant administration of 5-fluorouracil-based chemotherapy by continuous venous infusion or oral capecitabine, and who underwent pelvic MRI with a slice thickness of 3.5 mm both before and after nCRT, were included in the study. Inclusion criteria also involved the availability of the results of histopathological examination, complete with Mandard’s TRG classification on the surgical specimens.

Patients who did not meet these criteria, along with all patients with poor image quality and movement artifacts, were excluded. All the patients included gave written informed consent to use imaging data for research aims.

Image acquisition

Staging and restaging of all patients were performed using a 1.5T MRI (Avanto; Siemens, Erlangen, Germany) with phased-array coils. The examination protocol included T2-weighted sagittal, oblique coronal (i.e., parallel to the long axis of the primary tumor) and oblique axial (i.e., perpendicular to the long axis of the primary tumor) images, with a repetition time (TR) of 3790–5354 ms, an echo time (TE) of 95 ms, a slice thickness of 3.5 mm, a field of view (FoV) of 320 × 286 mm and an image matrix of 320 × 257.

Histopathological examination

Histopathological examination was performed by expert pathologists, specialized in gastrointestinal diseases. For each patient, a biopsy of the primary lesion before nCRT confirming the diagnosis of malignancy was available. Histopathologic examination on the surgical specimen included pathological tumor stage, nodal stage (in case of TME) and TRG according to Mandard’s classification, with TRG1 indicating complete response and TRG5 indicating the absence of regressive changes in the tumor [12]. Patients with complete response (TRG1), were classified as “group 0,” while patients with residual tumor (TRG2–5) were classified as “group 1.”

Image analysis

A volume of interest (VOI) was outlined around the rectal lesion in all paraxial T2-weighted MRI slices in which the cancer appeared, in both pre- and post-nCRT examinations. VOIs were obtained for all patients by two radiologists, who manually contoured the tumor area, using a dedicated software (PMOD, PMOD Group, Zurich, Switzerland). Radiologists were not aware of the final histopathological results but had access to pre-nCRT images when evaluating post-nCRT scans. Particular attention was paid to avoid endoluminal mucus and healthy rectal wall and, in post-nCRT MRI, to recognize parietal edema caused by the effects of nCRT (see Fig. 1). Texture analysis was performed using PMOD analysis algorithms; thirty-three parameters, derived from voxel histograms, gray-level co-occurrence matrix (GLCM) and run-length matrix (RLM) analysis, were extracted for each MRI scan.

Fig. 1
figure 1

a VOI manually drawn along the boundaries of a T3 rectal cancer in T2w image of staging MRI, excluding intraluminal mucus and normal rectal wall; b another VOI delineated around the tumor mass in a slice of a restaging MRI T2w, including both the lesion and fibrous tissue, excluding the inflammatory reaction in the nearby healthy mucosa

Full size image

Statistical analysis

Continuous variables obtained through TA were expressed as mean ± standard deviation (SD). All parameters derived from TA of both pre- and post-nCRT scans were dichotomized based on the results of histopathological examinations. The comparison between the two groups (responders, or “group 0” vs. non- responders, or “group 1”) was carried out using the Student's t test, with significance taken at p ≤ 0.05. For parameters with a statistically significant difference between the two groups, a receiver operating characteristic (ROC) curve was drawn and Youden index was calculated. Statistical analysis was performed using the R software [26].

Results

Patient clinicopathological characteristics

Sixty-two patients, 42 males (67.7%) and 20 females (32.3%), were included in this study. Patients clinicopathological characteristics are shown in Table 1. The majority of patients were older than 60 years (n = 42, 67.7%). Fifty patients (n = 50, 80.6%) underwent TME, while 12 patients (n = 12, 19.4%) underwent transanal local excision (LE). At histopathological examination, one third of patients achieved a pCR on the primary lesion (n = 20, 32.3%); 24 (38.7%) were T1–T2 and the remaining were T3–T4 (n = 18, 29.0%). Twelve patients (19.3%) had positive lymph nodes. Overall, 19 patients were defined as responses based on pTRG = 1 (30.6%), while 43 patients with pTRG between 2 and 5 (69.4%) were defined as non-responders.

Table 1 Patient clinicopathological characteristics
Full size table

Comparison of pre-nCRT T2-weighted TA parameters

Pre-nCRT TA parameters of the two groups are reported in Table 2. None of the parameters derived from the histogram, GLCM and RLM analyses showed a statistically significant difference between the two groups. The mean pre-nCRT skewness (1.04 ± 0.51 vs. 0.79 ± 0.31) and pre-nCRT excess kurtosis (2.28 ± 1.62 vs. 1.51 ± 0.91) were lower in group 1 compared to group 0; still, the difference did not meet statistical significance (p = 0.055 and 0.063 respectively).

Table 2 Comparison of pre-nCRT T2-weighted TA parameters using Student’s t test
Full size table

Comparison of post-nCRT T2-weighted TA parameters

Post-nCRT TA parameters for the two groups are reported in Table 3. Only GLCM maximum probability showed a statistically significant difference between the two groups (2909 ± 4479 vs. 6515 ± 8990; p = 0.039). Other parameters, such as GLCM homogeneity inverse different moment, GLCL sum entropy and GLCM entropy, approached the cut-off of p = 0.05, without reaching significance. GLCM maximum probability corresponds to occurrences of the most predominant pair of neighboring intensity values [27]. For this parameter, a ROC curve with an area under the curve (AUC) of 0.6622 (CI 95% 0.5128–0.8116) according to DeLong test was obtained [28]. The sensitivity and specificity at the Youden index were 14% and 100%, respectively (Fig. 2).

Table 3 Comparison of post-nCRT T2-weighted TA parameters using Student’s t test
Full size table
Fig. 2
figure 2

GLCM maximum probability ROC curve with Youden index

Full size image

Discussion

Management of LARC is constantly evolving. During the last decades, many efforts have been made in order to reach a tailored approach in the management of this disease, thus avoiding over-treatment and preserving patient’s quality of life. Non-operative management (NOM) of patients showing a cCR to nCRT can spare these patients the potential complications of major surgery, while other rectal preserving approaches, such as LE, can be used in some patients with a complete or major response in order to avoid the severe impact on quality of life of conventional surgery. On the other hand, neoadjuvant therapy itself is associated with several side effects, and is not always beneficial for the patients in terms of reduction of tumor burden.

Thus, the identification of reliable markers of response to nCRT has gained growing interest. TA has been suggested as a promising technique to investigate intra-tumor heterogeneity attributed to various factors such as hypoxia, necrosis and angiogenesis, potentially related to tumor aggressiveness and patient prognosis [20,21,22,23,24,25]. Both hypoxia and necrosis following chemo-radiotherapy cause fibrosis of the neoplastic tissue; fibrous tissue is characterized by hypo-intense signal in the T2-weighted sequences, and therefore can be differentiated from residual vital tumor [29]. TA could also predict which patients will respond to nCRT even before the treatment is started, or a few weeks after the onset of treatment [20, 21, 24, 25], and may improve MRI performance in the evaluation of response [22].

Different studies have investigated TA parameters of T2-wieghted MRI images as a tool to discriminate patients with a pCR and patients with residual disease. De Cecco et al. [20] found that only kurtosis obtained from T2-weighted pre-nCRT and mid-treatment MRI images showed significant differences between complete responders and partial or non-responders; the sensitivity and specificity for pCR prediction of the pre-nCRT T2-weighted kurtosis were 100% and 77.8%, respectively. In a later study of the same group, pretreatment T2-weighted images kurtosis showed a sensitivity and specificity with two different suggested cut-offs of, respectively, 100% and 67% and 83% and 83% [21]. Giannini et al. [23] investigated TA data in pre-nCRT MRI and 18F-FDG PET/CT images; only one second-order T2-weighted MRI texture parameter and five second-order PET texture parameters were included in a logistic regression model to differentiate responders from non-responders, with an area under the ROC curve of 0.86. One study, led by Shu et al. [24], identified multiple T2-weighted texture parameters as potential imaging biomarkers of response to therapy. The authors considered pre-, interim and post-nCRT T2-weighted TA, and found that skewness, entropy and energy of the pre-nCRT TA and variance, kurtosis, energy and entropy of early-nCRT TA (at 3 weeks from the beginning of nCRT) differed significantly between responders and non-responders [24]. Jalil et al. [22] also studied the application of T2-weighted TA in pre- and post-nCRT MRI scans, concluding that this technique can predict patients oncologic outcomes.

Nardone et al. [30] found in 49 patients a correlation between pre-nCRT MRI textural features and early disease progression. Yang et al. [31] found in 76 cases that first-order features of TA extracted from post-nCRT T2-weighted MR images could identify patients with complete response at histopathology. In 2019 and 2020, different authors proposed novel machine learning approaches based on MRI TA, obtaining promising results in the detection of complete responders to nCRT [32,33,34,35,36,37]. Some authors found correlations between MRI TA parameters and mutations status in rectal cancer [38,39,40] and nodal invasion [41].

The current study tested the efficacy in predicting response to nCRT with TA in 62 patients treated at the same third-level center and restaged with the same MRI scanner with the same protocol, thus conferring a good consistency to our findings. Our results do not support the routinary use of T2-weighted images TA for the evaluation of rectal cancer response to chemoradiation, even if one parameter obtained from MRI images after nCRT showed to be a potential imaging bio-marker to identify patients with a pCR. In fact, among the multiple parameters obtained in both pre- and post-nCRT images, only one, i.e., GLCM maximum probability post-nCRT, showed significant differences between complete responders (i.e., group 0 = TRG1) and non-responders (i.e., group 1 = TRG2–5). This parameter showed a low sensitivity, although a specificity of 100% was found at the ROC curve Youden index.

However, T2-weighted images TA obtained a worse accuracy compared to mrTRG. In the literature, mrTRG1-2 (i.e., complete or very good radiological regression) showed a sensitivity and specificity for predicting cCR of 74.4% and 62.8% respectively [29], and a good correlation to long-term oncologic outcomes [16]. Finally, in our experience, TA proved to be time-spending for the operator, who must use a special software to accurately delineate the tumor circumference in all the MRI slices.

There are several limitations that need to be considered when interpreting our data. The number of patients included in our study is undoubtedly not large enough to draw definitive conclusions, even though many previous studies on MRI TA of rectal cancer have investigated a lower or similar number of cases [20,21,22,23, 30, 31]. Also, the retrospective design of the study and the different technique used for MRI examinations (endorectal gel vs. air rectal distension) should be taken into account as potential confounding factors.

In conclusion, we did not find any strong evidence for the use TA as a reliable tool for differentiating TRG1 patients from patients with residual disease. Based on our findings and taking into consideration the complexity of the analysis, we consider MRI T2w texture parameters neither suitable to cover the role of predictors of the histopathological response of rectal cancer patients to neoadjuvant therapy, nor to substitute mrTRG evaluation.