Early prediction of pathological response in locally advanced rectal cancer based on sequential ¹⁸F-FDG PET

Abstract

Background

The objectives of this study were to investigate the predictive value of sequential ¹⁸F-FDG PET scans for pathological tumor response grade (TRG) after preoperative chemoradiotherapy (PCRT) in locally advanced rectal cancer (LARC) and the impact of partial volume effects correction (PVC).

Methods

Twenty-eight LARC patients were included. Responders and non-responders status were determined in histopathology. PET indices [SUV max and mean, volume and total lesion glycolysis (TLG)] at baseline and their evolution after one and two weeks of PCRT were extracted by delineation of the PET images, with or without PVC. Their predictive value was investigated using Mann-Whitney-U tests and ROC analysis.

Results

Within baseline parameters, only SUV_mean was correlated with response. No evolution after one week was predictive of the response, whereas after two weeks all the parameters except volume were, the best prediction being obtained with TLG (AUC 0.79, sensitivity 63%, specificity 92%). PVC had no significant impact on these results.

Conclusion

Several PET indices at baseline and their evolution after two weeks of PCRT are good predictors of response in LARC, with or without PVC, whereas results after one week are suboptimal. Best predictor was TLG reduction after two weeks, although baseline SUV_mean had smaller but similar predictive power.

Preoperative radiochemotherapy (PRCT) is now considered a standard treatment for patients diagnosed with locally advanced rectal cancer (LARC). A significant tumor downsizing and downstaging, as well as a reduction of the risk for local recurrence and longer survival have been observed as a result of PRCT [1–3]. Within this context, ¹⁸F-FDG PET imaging has been demonstrated as an interesting predictive tool [4]. Indeed, correlations between the pathological tumor response after PRCT and the standard uptake value (SUV) decrease within the tumor have been demonstrated in several studies [5–9]. In most of these studies, the SUV evolution between the baseline positron emission scan (PET) scan and the one acquired before surgery was correlated with the tumor regression grade (TRG) [5–9]. However, allowing for earlier prediction of the response is of higher interest for the clinical practice, since it might enable modifications of the treatment protocol [10]. A few studies have investigated the early prediction of the pathological tumor response based on ¹⁸F-FDG PET imaging during PRCT [8,9]. Cascini et al. showed that changes in the metabolic activity of the tumor, measured as early as 15 days after the start of PRCT, were predictive for response [8]. More recently, Janssen et al. have examined three different time points of ¹⁸ F-FDG PET imaging during PRCT, in order to define the optimal time for early prediction [9]. In this study ¹⁸F-FDG PET scans at baseline, as well as at 8 and 15 days during PRCT were carried out and it was found that the best predictive factor of TRG was the SUV_max response index (RI, defined as the percent evolution relative to the pretreatment value) at 15 days. All these studies have considered SUV measurements (max and/or mean) only. It has been demonstrated recently in several studies and various malignancies that other ¹⁸ F-FDG PET derived parameters more fully characterizing tumors on a functional level can have statistically higher predictive value than SUV [10–17]. These include metabolically active tumor volume (MATV, defined as the tumor volume as it can be seen and delineated on a PET image) [11] and total lesion glycolysis (TLG, multiplying MATV and its associated mean SUV) [12]. In addition, some of these studies demonstrated that response could be predicted by extracting these parameters from the pretreatment scan only, therefore potentially eliminating or reducing the need for sequential scans during treatment. Such results have been presented within the context of locally advanced esophageal cancer (LAEC) [11,13], non-Hodgkin lymphoma [14], pleural mesothelioma [15] and cervix and head and neck cancers [16], whereas conflicting results have been recently obtained in rectal cancer [17,18]. Finally, most of these studies have considered the evolution of PET derived parameters without partial volume effects (PVE) correction (PVC). This may lead to biased results especially if MATVs change size and/or shape during treatment. The current study was therefore conducted retrospectively on the cohort previously imaged [9] with the following objectives: 1) determine the predictive value of baseline ¹⁸F-FDG PET derived parameters; 2) investigate the evolution of these parameters during treatment and their associated predictive value; and 3) investigate the impact of PVC on these results.

Material and methods

Patients

This study consisted of a retrospective analysis of a cohort of 28 patients (see Table I) diagnosed with non-metastasized LARC previously recruited in a prospective imaging study [9] approved by the medical ethics committee and for which all patients gave written informed consent before entering the study. Although all tumors may not strictly speaking be considered as locally advanced according to staging, none of the patients could be resected immediately because of the high risks of differentiation, bulkiness, and size and location of the tumor. They were therefore all treated with the same PCRT protocol as recommended for LARC patients. All patients underwent PRCT (28 fractions of 1.8 Gy, 5 fractions a week, and concomitant capecitabine, 825 mg/m², twice a day), followed by surgery (total mesorectal excision). Patients underwent ¹⁸F-FDG PET/CT scans at baseline, and on Days 8 and 15 of treatment. From here onwards, these scans will be denoted as PET_i, with i from 1 to 3.

Table I.

Patients’ characteristics.

Parameter	Number of patients (%)
Gender
Male	18 (64)
Female	10 (36)
Age
Range	44–81
Median	67
TNM Stage
T1	0 (0)
T2	1 (4)
T3	24 (85)
T4	3 (11)
N0	4 (14)
N1	8 (29)
N2	16 (57)
M0	28 (100)
M1	0 (0)
TRG status
1	6 (21)
2	6 (21)
3	9 (33)
4	6 (21)
5	1 (4)

¹⁸F-FDG PET/CT acquisitions

The protocol was designed to ensure robust SUV measurements across all three time points. Patients received an intravenous injection of FDG after a minimal fasting period of 6 h, with the activity normalized for the weight of the patient as follows:

$$\left(\text{weight}[\text{kg}]\times 4+20\right)[\text{MBq}]$$

After an uptake period of exactly 60 min, all acquisitions were carried out on a Siemens Biograph 40 TruePoint scanner (Siemens Medical, Erlangen, Germany) with a spatial resolution of approximately 6 mm at full-width-at-half-maximum. Listmode data of the abdominal region were acquired in three-dimensional (3D) mode with scatter, decay, and computed tomography (CT)-based attenuation corrections, and 5 min per bed position. Images were reconstructed using Fourier rebinning and 2D OSEM (four iterations, eight subsets) with voxel size 4.1 × 4.1 × 3 mm³ without post-filtering.

Pathological tumor response grade (TRG)

The ground-truth of tumor response to therapy was determined in histology since all patients underwent total mesorectal excision. As proposed by Mandard et al. [19], TRG for each tumor was determined by an experienced pathologist blinded to the imaging data as follows: TRG 1, complete tumor response; TRG 2, residual cancer cells scattered through fibrosis; TRG 3, an increased number of residual cancer cells, with predominant fibrosis; TRG 4, residual cancer outgrowing fibrosis; and TRG 5, no regressive changes within the tumor. Tumors were subsequently grouped into responders (TRG 1–2) and non-responders (TRG 3–5).

Partial volume effects correction

All PET images were corrected for PVE using an iterative deconvolution method previously validated on simulated and clinical datasets [20]. This approach iteratively estimates the corrected voxels values through Lucy-Richardson deconvolution [21,22] with prior knowledge (within ± 1 mm, as it has been shown that a 1 mm error in the PSF led to a negligible impact on measured SUVs [23]) of the scanner’ s Point Spread Function (PSF), assumed to be spatially invariant in the field of view. In this study the tumors were all in the exact same body region and this assumption has therefore no significant impact on the applied correction on a patient-by-patient comparison basis. Wavelet-based denoising was incorporated using Bayeshrink filtering [24], applied to the residual within each iteration of the deconvolution process. This allows using a sufficient number of iterations to correct for PVE without significant noise addition. This methodology is voxel-based and therefore does not assume homogeneous regional radiotracer distributions for the tumor and/or surrounding background.

Investigated parameters and analysis

All parameters were extracted from the original (PET_1–3 and PVE corrected (PET_1–3^PVC) images, denoted from here onwards param_1–3 and param_1–3^PVC respectively. For each patient, the tumor was identified on each PET i images by a nuclear physician with more than 10 years experience and subsequently semi-automatically isolated from the bladder in a 3D region of interest (ROI) using an in-house software. This ROI, containing only the MATV and its surrounding background, was automatically transferred to the corresponding corrected PET_i^PVC image. MATV were subsequently delineated on both uncorrected and corrected images using an implementation of the Fuzzy Locally Adaptive Bayesian (FLAB) automatic algorithm [25] in the same software. The FLAB approach allows automatic tumor delineation by computing a probability of belonging to a given “class”, e.g. tumor or background, for each voxel within the 3D ROI. This probability is calculated by taking into account the voxel’ s intensity with respect to the statistical distributions (characterized by their mean and variance) of the voxels in the various regions of the image, as well as its spatial correlation with neighboring voxels in 3D. This approach has been validated on simulated and clinical datasets for accuracy, robustness and reproducibility, on both homogeneous and heterogeneous MATVs [25–27].

SUV_max, and SUV_mean as well as the MATV and the TLG were automatically calculated from these delineations. MATV was defined as the sum of all voxels contained in the FLAB delineated volumes multiplied by the volume of a voxel (50.43 mm³). TLG was determined by multiplying the MATV and its associated SUV_mean. Response indices (RIs) corresponding to one (RI₂) and two (RI₃) weeks were calculated as the percentage evolution with respect to the baseline value (PET₁) as follows: RI_n (param_n−param₁)/param₁ × 100 for n = 2 and 3. Similarly, RI_2–3^PVC were calculated using the parameters extracted from PET_1–3^PVC images.

Figure 1 illustrates for a non-responder (Figure 1A) and a responder (Figure 1B) the baseline scan and the scan after two weeks, with the delineation of the tumor on both scans.

Figure 1.

Illustration of the baseline scan (on the left) and the scan at 2 weeks (PET3, on the right) for (A) a responder and (B) a non-responder. The green contour is the FLAB delineation.

Statistical analysis

Statistical analyses were performed using Medcalc™ (MedCalc Software, Belgium). All quantitative values were expressed as means ± standard deviations (SD) and ranges (minimum–maximum). Binary response status based on TRG classification (1–2 vs. 3–5) was correlated with baseline values (param₁) as well as RI_2–3 for early sequential scans (PET_2–3) using a Mann-Whitney U-test. For the parameters that were found to be significantly correlated with response, the predictive performance regarding the identification of non-responders was evaluated using ROC analysis. Area under the curve (AUC), along with the best compromise between sensitivity and specificity were reported, as well as results associated with a specificity of 100%, corresponding to the clinical goal of identifying non-responders without erroneously identifying any of the responders. For RI_2–3, only cut-off values above +30% or below −30% were considered. This constraint was applied because the reproducibility (or test-retest assessment) of the PET derived parameters (both SUV and volume-based measurements) under investigation here has been previously determined to be at such upper and lower limits [27]. Thus characterization of response (or disease progression) based on evolution of PET derived indices must take into account such reproducibility limits. All tests were two-sided and p-values below 0.05 were considered statistically significant.

Results

Statistics for baseline and evolution of all derived parameters (with or without PVC) are given in Table II. Results of Mann-Whitney U-tests and subsequent ROCs analyses are provided in Supplementary Table I online at http://www.informahealthcare.com/doi/abs/10.3109/0284186X.2012.702923 and Table III, respectively.

Table II.

Baseline values and evolution (RIs) of each PET derived parameter with respect to baseline values.

Parameter	Baseline value median ± SD (mean, min, max)	RI2 (%) median ± SD (mean, min, max)	RI3 (%) median ± SD (mean, min, max)
SUVmax	13.9 ± 5.2 (14.9, 6.6, 26.7)	−18 ± 39 (−12, −63, +155)	−32 ± 25 (−31, −73, +47)
SUVmaxPVC	21.2 ± 9.7 (23.8, 10.2, 49.3)	−20 ± 41 (−15, −59, +165)	−43 ± 27 (−38, −75, +47)
SUVmean	7.8 ± 2.8 (8.0, 3.7, 13.9)	−20 ± 36 (−11, −59, +138)	−29 ± 25 (−28, −66, +54)
SUVmeanPVC	10.2 ± 4.0 (10.3, 4.5, 20.8)	−22 ± 37 (−12, −56, +153)	−28 ± 24 (−29, −67, +42)
MATV	23 ± 79 (49, 2, 397)	−18 ± 22 (−15, −51, +51)	−36 ± 29 (−32, −80, +42)
MATVPVC	19 ± 76 (45, 2, 378)	−15 ± 22 (−15, −51, +39)	−32 ± 28 (−32, −78, +37)
TLG	181 ± 473 (345, 10, 2235)	−35 ±45 (−22, −75, +137)	−57 ± 37 (−48, −86, +68)
TLGPVC	196 ± 518 (379, 13, 2385)	−34 ± 41 (−24, −75, +130)	−59 ± 33 (−50, −88, +51)

Table III.

Predictive value of each parameter baseline values and RIs using ROCs analysis. Only parameters for which a significant correlation was found [see appendix, to be found online, at http://www.informahealthcare.com/doi/abs/10.3109/0284186X.2012.702923] are included.

Parameter*		AUC	Cut-off value	Sensitivity (%)	Specificity (%)	Sensitivity for 100% specificity (%)	Cut-off value
SUVmax	RI3 (%)	0.77	−43	88	58	31	−16
SUVmaxPVC	RI3 (%)	0.73	−33	50	83	13	−5
SUVmean	PET1	0.75	6.1	50	100	50	6.1
	RI2 (%)	0.73	−31	88	33	50	−10
	RI3 (%)	0.79	−34	81	67	31	−11
SUVmeanPVC	PET1	0.76	7.3	44	100	44	7.3
	RI2 (%)	0.77	−33	81	25	50	−8
	RI3 (%)	0.75	−42	88	58	19	−8
TLG	RI2 (%)	0.75	−34	56	67	56	−23
	RI3 (%)	0.79	−53	63	92	38	−37
TLGPVC	RI2 (%)	0.70	−34	56	50	44	−18
	RI3 (%)	0.78	−54	63	92	31	−38

Pathologic tumor response

Six tumors were characterized by complete pathologic response (TRG 1) and six had residual cancer cells (TRG 2). Nine, six and one tumors were classified as TRG 3, 4 and 5, respectively (Table I). There were therefore 12 responders (43%) and 16 non-responders (57%).

Predictive value of baseline parameters

Without PVC

Pretreatment SUV measurements were all normally distributed, with SUV_max1 and SUV_mean1 values for the cohort of 14.9 ± 5.2 and 8.0 ± 2.8, respectively. On the other hand, MATV₁ and TLG₁ were not normally distributed. Median MATV₁ and TLG1 values were 23 cm³ and 181 g, respectively. Most volumes were within the range of 10–40 cm³, with five tumors above 100 cm³.

Responding tumors were characterized by higher baseline SUV (SUV_max1 15.6 ± 3.7 and SUV mean 9.0 ± 2.3) than non-responding ones (SUV_max1 12.5 ± 5.8 and SUV_mean 6.4 ± 2.9). The difference was not significant for SUV_max (p 0.06) contrary to SUV_mean (p = 0.02), with an associated AUC of 0.75 (Table III) leading to perfect specificity (100%) but an associated sensitivity of 50% for a cut-off value of 6.1.

No correlation between response and MATV₁ or TLG₁ was found (p > 0.1), distributions of these parameters among responders and non-responders being largely overlapped (Table II).

With PVC

Pretreatment SUV_max and SUV_mean increased by + 57 ± 23% and + 27 ± 10% after correction, to 21.2 ± 9.7 and 10.2 ± 4.0, respectively. On the other hand, MATV₁^PVC were systematically smaller with however a small difference of only −12 ± 6% (range −3% to −24%). Resulting from a combination of smaller MATV and higher SUV_mean, TLG₁^PVC increased by +11 ± 7% (range 1% to +31%). The increase of SUV_mean was strongly correlated with MATV (r > 0.7, p <.0001) whereas increase of SUV max was not (r <.5). Despite these absolute values increase, the differences between responding and non-responding tumors were not significantly altered by PVC. SUV_max1^PVC still did not significantly separate responding from non-responding tumors (p = 0.07) whereas SUV_mean1^PVC allowed similar differentiation (p =.02), and no correlation with response was found for MATV₁^PVC and TLG₁^PVC (p >.2).

Regarding ROC analysis, AUCs and associated optimal sensitivities/specificities with PVC were mostly similar as without correction. For SUV_mean1^PVC, AUC increased from 0.75 to 0.76, although it led to a reduced sensitivity (44% instead of 50%) with the same specificity of 100% (Figure 2A).

Figure 2.

ROCs examples for prediction of non-responders (n = 16), using (A) SUV_mean with or without PVC, (B) TLG and SUV_mean RI₃, and (C) logistic regression combining baseline MATV and SUV_mean values and their evolution after 8 and 15 days, with and without PVC.

Evolution of parameters during PRCT and associated predictive value

Without PVC

There was a global trend to decreasing MATV and associated uptake of the tumors during PRCT across the cohort of patients, with further reduction at 15 days (Table II). However, these PET parameters were also found to be increasing for some patients (Table II), especially after eight days. The lowest decreases at eight days were observed for MATV and SUV_max (mean −12% and −15%, median −18% for both) whereas the largest decrease was observed for TLG with a mean of −22% (median −35%). Decrease was systematically larger at 15 days, with for instance −48 mean TLG decrease (median −57%).

On the one hand, none of the RI₂s were statistically different between responders and non-responders, except for SUV_mean (−6 ± 44% vs. −25 ± 12%, p =.04) and TLG (−18 ± 53% vs. −37 ± 15%, p =.03). On the other hand, RI₃s of all the parameters except MATV were statistically different between responders and non-responders, especially for TLG and SUV_mean (p = 0.009).

According to ROCs analysis, RI₂s of neither SUV_mean nor TLG (the only two parameters for which there was a statistical difference at eight days between responders and non-responders) allowed satisfactory prediction of TRG. With an AUC of 0.73, the optimal cut-off value above 30% change (−31%) for SUV_mean was associated with a sensitivity of 88% but a specificity of only 33%. RI₂ of TLG led to an AUC of 0.75 and much lower sensitivity (56%) but higher specificity (67%) with a cut-off value of −34%. At eight days, 100% specificity could be achieved only using too low cut-off values (−10% only for SUV_mean, and −23% for TLG) and at the cost of reduced sensitivity for SUV_mean (50%).

Predictive performance improved at 15 days with higher sensitivities/specificities for all parameters. For SUV_max, a −43% RI₃ cut-off value was associated with 88% sensitivity and 58% specificity, whereas a −34% RI₃ for SUV_mean led to 81% sensitivity and 67% specificity. Reduction of TLG using a RI₃ cut-off value of −53% led to lower sensitivity (63%) but higher specificity (92%) (Figure 2B). At 15 days, maximizing specificity was associated with low RI₃ cut-off values (−11 and −16% for SUV_mean and SUV_max), except for TLG for which the cut-off value was −37%, with however a sensitivity of only 38%. Sensitivities associated with 100% specificity were also between 30% and 40% for the other parameters.

With PVC

Evolution of the parameters after 8 and 15 days was not significantly altered by PVC, with very similar values of RIs for all parameters (see Table II). PVC had therefore no impact on the statistical difference between RIs of responders and non-responders. After correction, the RI₂s still did not allow statistical differentiation between responders and non-responders for most parameters, except for TLG and SUV_mean which was already the case without PVC. Similarly, all the RI₃s except MATV were still able to statistically differentiate the non-responders from the responders after PVC, but the differentiation was significantly neither improved nor reduced.

Similarly, ROCs results were not significantly altered by PVC, although AUCs varied by up to ± 0.05. Sensitivities and specificities corresponding to optimal cut-off values on the other hand were in a few cases significantly modified. For example, the sensitivity associated with RI₃ of SUV_max was decreased after PVC from 88 to 50%, with an increase of specificity from 58 to 83%. Similar changes were observed for the other parameters at both 8 and 15 days, without nonetheless significantly reduce or increase the predictive performance of the parameters.

Discussion

Neither the predictive value of baseline PET images nor the one of PET features more fully describing tumors (MATV, TLG) has been extensively determined yet within the context of PCRT early prediction in rectal cancer. Melton and colleagues found in 21 LARC patients that the reduction of PET-based parameters (MATV, TLG and SUV) between baseline and four to six weeks after treatment was correlated with pathological response [17], but they did not investigate the value of baseline parameters. More recently, Chennupati and colleagues did not find a correlation between pathological response and SUV_max or MATV measurements at baseline or reduction after treatment in a cohort of 35 LARC patients, but they did not investigate the value of parameters reduction during the treatment [18]. Janssen and colleagues demonstrated that the reduction of SUV max after two weeks was a good predictor of response, but they did not investigate other PET parameters beyond SUV_mean or their baseline predictive value [9]. None of these studies investigated the impact of PVC. To the best of our knowledge, the current study is the first investigation on a homogeneous LARC cohort including the impact of PVC, four PET derived parameters (SUV_max, SUV_mean, MATV, TLG), and three (baseline and after one and two weeks) ¹⁸F-FDG PET scans during PCRT.

On the one hand, contrary to results in LAEC [10], neither the baseline MATV nor the derived TLG were significantly associated with response. On the other hand, some of the baseline parameters were of predictive value, such as higher SUV_mean being associated with responding tumors (p =.02), and allowed prediction of the non-responders with a specificity of 100% but a limited sensitivity of 50%. A similar trend was observed with SUV_max, although without reaching statistical significance (p =.06).

Overall, response was associated with higher decrease of the ¹⁸F-FDG PET derived indices within the first two weeks of PCRT, especially SUV_mean and TLG (p <.01), although this decrease was a more efficient predictive factor after two weeks than one, in line with previous findings [9]. First, none of the parameters’ RI₂s except for SUV_mean and TLG were significantly correlated with TRG. Second, an important constraint was that cut-off values of RI_2–3 s had to be larger than ± 30%, because of the upper and lower reproducibility limits of such PET derived measurements [27]. However, measured RI_2–3 s were often within this range, especially after one week, although it was still the case even after two weeks. This may constitute one of the major limitations of early PET based response prediction, and the reproducibility of PET acquisitions clearly need to be improved in that regard. On the one hand, according to ROCs analysis, the reduction of SUV_mean after two weeks was associated with an AUC of 0.79, a 100% specificity and 31% sensitivity associated with a RI₂ cut-off value of −11%, which is too low with regard to the reproducibility limits, as it is the case for all other parameters. The only exception was TLG: a −37% RI₃ cut-off value led to 100% specificity and 38% sensitivity, whereas a −53% cut-off value led to the best compromise of 63% sensitivity and 92% specificity.

Usually, higher SUVs are associated with more aggressive tumors, and therefore may be potentially more resistant. However according to our results, responding tumors were found to have higher initial SUVs. This can be associated with the fact that they were also the ones exhibiting the highest uptake decrease during treatment. In addition, the tumors that were classified as non-responders based on the RI₃ of TLG (specificity 100%, sensitivity 38%) are almost the same as those classified as non-responders based on their baseline mean SUVs (specificity 100%, sensitivity 50%). Therefore waiting for two weeks did not significantly (p > 0.05) improve prediction of response in this context when maximizing specificity, despite increased overall AUCs, which is a new result with respect to previous findings that only investigated the predictive value of RIs and not absolute baseline values [9].

The classification could be further improved using logistic regression of several parameters such as baseline MATV and SUV_mean values and their evolution at 8 and 15 days, resulting in an AUC of 0.88, with a sensitivity and specificity of 88% and 83% respectively, whereas a specificity of 100% was associated with 50% sensitivity (Figure 2C). Similar results were obtained with PVC (AUC 0.92, 92% specificity and 81% sensitivity, 56% sensitivity associated with 100% specificity). However, such weighted model combining more than one parameter on 28 patients only is likely to be over fitted to the data and would require validation on a larger cohort.

Regarding PVC, despite significant impact on the absolute values (except for MATV), it had limited impact on the predictive value of the parameters. Similar results regarding the impact of PVC on baseline predictive value were obtained for baseline prediction of chemoradiotherapy response in LAEC [28]. This is the first study however investigating the PVC impact on the predictive value of sequential scans. This lack of impact may be explained by the fact that although the tumor volumes significantly shrunk during the first two weeks of PCRT (−36 ± 29% after two weeks), their volume were still large (mean 36 cm³). Since PVE are significant for volumes below 10–15 cm³, PVC impact should be investigated for the follow-up of tumors exhibiting larger changes and/or smaller volumes.

Some limits of this study have to be emphasized. It is first limited by its retrospective nature and the small number of patients, which led to group the five different classifications of TRG into responders and non-responders. Investigation of other response classification such as complete regression (TRG1) vs. non-response (TRG 2–5) or all five TRG statuses should be conducted in larger prospective studies in order to validate our preliminary results. Finally, the TRG classification, although it constitutes the best measurement of tumor response available, was determined by one pathologist only.

Conclusion

Our results in this preliminary study suggest that early prediction of tumor response to PRCT in rectal cancer using sequential ¹⁸F-FDG PET scans is challenging due to limited reproducibility of PET imaging, especially after one week that was found to be suboptimal in both robustness due to small evolutions of the parameters and accuracy with no correlation with response. The best compromise between accuracy and robustness of prediction was obtained by considering the reduction of total lesion glycolysis after two weeks of PCRT, with a sensitivity of 63% and specificity of 92%, although the improvement over baseline absolute mean SUV prediction (specificity 100%, sensitivity 50%) was not significant. Finally, partial volume effects correction had no impact on the predictive value of neither the baseline absolute values nor their evolution during treatment. The results of this study require validation in a larger cohort allowing consideration of less restrictive response measures.