This prospective study had approval from our hospital’s ethics review committee and each patient provided signed informed consent. From July 2022 to July 2023, 88 individuals with substantial clinical indications of rectal cancer were submitted to pelvic MRI. The following patients were excluded: (1) histological results indicating non-rectal cancer (n = 5); (2) missing clinical or histological information (n = 5); (3) poor image quality for glucoCEST, APTWI, or DWI (n = 8); (4) inability to complete all imaging sequences due to claustrophobia or other physical symptoms (n = 2); (5) previous radiotherapy, chemotherapy, or surgery prior to MRI (n = 8). Ultimately, a total of 60 patients were included, and clinical data, including gender, age, maximum lesion diameter, classification, location, and tumor stage, were collected. The study flowchart is shown in Fig. 1.

MRI was obtained with a 16-channel phased array torso coil on a 3.0-T MR scanner (Ingenia 3.0 T, Philips Healthcare, Best, the Netherlands). Prior to the examination, all patients were instructed to evacuate the rectum and were administered 20 mg of scopolamine butyl bromide (Buscopan; Boehringer) intramuscularly to reduce gastrointestinal movements. Subsequently, each patient was placed supine, feet first, and a respiratory belt was fastened around the abdomen to monitor respiration. DWI (b = 0, 1000 s/mm2), APTWI, and glucoCEST were performed, followed by T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and enhanced T1WI (Magnevist, 0.2 mL/kg body weight, Bayer Healthcare). APTWI was performed with a 3D single-shot TSE-Dixon pulse sequence with inversion recovery (SPIR) fat suppression. The duration of saturation pulse train for the APTWI sequence was 1 s, with a B1 root mean square (B1, rms) of 1.0 μT. Seven saturation frequency offsets (± 3.5, ± 3.42, ± 3.58, −1560 ppm) were applied to acquire APTWI images. At +3.5 ppm, three images with different echo times (ΔTE = 0.4 ms) were acquired for the asymmetry calculation of B0 map applying the mDixon algorithm. GlucoCEST was performed with a 2D single-shot TSE-mDixon sequence with SPIR fat suppression. The saturation frequency and duration of radiofrequency saturation pulse train for the glucoCEST sequence were 1.2 ppm and 1 s, respectively, and B1 root mean square (B1, rms) was 2.0 μT. To acquire a z-spectrum, imaging was repeated at 20 saturation frequency offsets of −5, −3.5, −3, −2, −1.5, −1.2, −0.8, −0.5, −0.25, 0, 0.25, 0.5, 0.8, 1.2, 1.5, 2, 3, 3.5, 5 ppm, as well as one offset far from the resonant frequency (1024.0 ppm). The protocol’s details are provided in Table 1.

All images were uploaded to the post-processing Intellispace Portal Workstation (ISP, version 9, Philips Healthcare, the Netherlands), to generate the ADC map with the MR diffusion software package. B0/B1 mapping and B0-corrected 3D APTWI images were auto-constructed on the MR scanner. 2D-glucoCEST images were post-processed with a custom-made program in MATLAB software for the acquired z-spectra and MTRasym (1.2 ppm) mapping [21]. The DWI-derived parameter apparent diffusion coefficient (ADC) was calculated using Eq. 1:

$$_}}}}=}}}}_\cdot \exp (-}}}\cdot ADC)$$

(1)

where b represents the diffusion sensitizing factor; S0 and Sb represent signal intensities (SIs) at a b-value of 0 and the b-value indicated by the subscript, respectively. The glucoCEST- and APTWI-derived parameters magnetization transfer ratio asymmetry at 1.2 ppm downfield from the water signal (MTRasym (1.2 ppm)) and magnetization transfer ratio asymmetry at 3.5 ppm downfield from the water signal (MTRasym (3.5 ppm)) were derived from Eqs. 2 and 3, respectively:

$$MTR}}}(3.5\,ppm)=[_(-3.5\,ppm)-_(+3.5\,ppm)]/_$$

(2)

$$MTR}}}(1.2\;ppm)=[_(-1.2\;ppm)-_(+1.2\;ppm)]/_$$

(3)

where S0 and Ssat are signal intensities (SIs) obtained without and with selective saturation, respectively. The regions of interest (ROIs) were manually placed along the mass tumor area at the slice of the largest area on axial DWI images with reference to T2WI and enhanced T1WI images. Areas with cystic degeneration, necrosis, apparent signs, hemorrhage artifacts and blood vessels were avoided. Then, the ROIs were copied to the ADC, MTRasym (1.2 ppm), and MTRasym (3.5 ppm) pseudo-color maps, and mean ADC, MTRasym (1.2 ppm), and MTRasym (3.5 ppm) values of the ROIs were automatically determined. All the above procedures were independently performed by an attending radiologist and an associate chief radiologist with 8 and 15 years of experience in pelvic MRI, respectively, who were blinded to each other’s results and to the clinicopathological data of patients.

All specimens were obtained by surgery within 2 weeks of the MRI examination and sent to the pathology department of our hospital for analysis by a pathologist with 10 years of experience in gastrointestinal pathologic diagnosis who was blinded to imaging data. Hematoxylin-eosin (H&E) staining was performed to determine histological features. Referring to gland or tubule formation and architecture in the lesion, each rectal cancer specimen was categorized as grade 1 (well-differentiated, > 95% gland formation), grade 2 (moderately differentiated, 50–95% gland formation), and grade 3 (poorly differentiated, < 50% gland formation) [22]. Grade 1 and 2 samples were assigned to the low-grade group, while grade 3 cases were in the high-grade group.

R (version 3.5.3; R Foundation, Auckland, New Zealand) and SPSS (version 15.0; MedCalc Software, Ostend, Belgium) were used for data analysis. Interobserver agreement between the two radiologists for glucoCEST SIs, MTRasym (3.5 ppm), and ADC values was assessed using the intra/interclass correlation coefficient (ICC), with an ICC > 0.75 indicating adequate reliability [13]. Between the high- and low-grade rectal cancer groups, categorical variables were compared by the chi-square test, while continuous variables were compared by the Mann–Whitney U test or independent samples t-test. The diagnostic efficacy was evaluated using the area under the receiver operating characteristic curve (AUC), and AUCs were compared by the DeLong test. Logistic regression (LR) analysis was employed to identify independent predictors and generate a multi-parameter composite diagnostic tool. A control model was built by bootstrapping (1000 samples), and its performance was assessed by calibration curve analysis, decision curve analysis (DCA), and ROC curve analysis. Statistical significance was set at p < 0.05.