Colorectal Cancer (CRC) is regarded as the third most common cancer and a leading cause of cancer-related deaths globally. Rectal cancer constitutes 30-35% of the CRC cases and approximately 30-40% diagnosed at an advanced stage [1,2].   Locally Advanced Rectal Cancer (LARC) (T3-4 or node-positive) is usually treated with neoadjuvant chemoradiotherapy (nCRT) followed by surgery [3]. While nCRT reduces recurrence, downstages tumors and increases sphincter-preservation rates [4], only 4-31% of patients achieve pathological complete response (pCR), who may benefit from organ-preserving strategies like local excision or watch-and-wait management [5,6].   Furthermore, MRI is the preferred modality for rectal cancer staging and restaging. It plays a major role in assessment of treatment response, monitoring of recurrence and surgical planning assistance [7].   Meanwhile, Diffusion-Weighted Imaging (DWI)/Apparent Diffusion Coefficient (ADC) reflects tissue cellularity, enabling differentiation between residual tumor, fibrosis and necrosis post-treatment. The ADC values and changes assess nCRT response in LARC [8].   Also, radiomics emerged as a promising tool for predicting prognosis and guiding treatment decisions through high-throughput extraction of quantitative imaging features from medical images [9]. Texture analysis, a subset of radiomics, measures intratumoral heterogeneity. The T2-weighted and ADC-based texture analysis, including first-order histogram features-entropy, skewness, kurtosis-assist as potential biomarkers for predicting treatment response in rectal cancer [10-13].   Most studies analyzed a limited number of texture parameters without integrating them into predictive models or validating their diagnostic performance against established metrics such as MRI TRG. To fill this gap, this study evaluates the combined diagnostic performance of ADC-based radiomic texture features and MRI TRG for predicting pCR following nCRT in LARC. What distinguishes the present study from prior radiomics research is threefold. First, while previous ADC texture analyses have predominantly evaluated individual histogram parameters (e.g., skewness, kurtosis) in isolation, we develop and compare three hierarchical predictive models: MRI TRG alone, MRI TRG combined with mean ADC and a comprehensive ADC radiomics model incorporating multiple first-order features. Second, unlike most published studies that report radiomic performance without benchmarking against established clinical imaging metrics, we directly compare the incremental value of adding ADC texture features to MRI TRG-a standardized clinical reporting tool. Third, whereas prior work has largely focused on pre-treatment ADC features for response prediction, our analysis integrates both pre- and post-treatment delta (Δ) parameters (Δmean ADC, Δskewness, Δkurtosis, Δentropy) to capture dynamic treatment-induced microstructural changes. This combined approach-integrating morphological (mrTRG), functional (ADC values) and texture-based (histogram features) parameters within a comparative modeling framework-has not, to our knowledge, been previously reported.

We conducted a prospective cohort study at a single center of Rzgary teaching Hospital, including 43 consecutive patients diagnosed with LARC between November 2024 and November 2025. All patients had histopathologically confirmed adenocarcinoma.   No formal sample size or power calculation was performed before study initiation, as this was an exploratory prospective study without prior effect size estimates for ADC radiomic parameters in the local population. The sample size (n = 34) was determined by the number of eligible LARC patients who met the inclusion criteria during the 13-month study period (November 2024 to November 2025).   pCR Event Rate and Model Stability Considerations Among the 34 included patients, 9 (26.5%) achieved pathological complete response. This event rate is consistent with published literature (4-31%) [5,6]. However, for binary logistic regression and ROC-based predictive modeling, a minimum of 10-20 Events per Predictor Variable (EPV) is recommended to reduce overfitting and ensure stable parameter estimates. With only 9 pCR events, the present study does not meet this EPV threshold, particularly for the full ADC radiomics model incorporating multiple first-order texture features. Consequently, the reported performance metrics-especially the AUC of 1.000-are likely optimistic and should be interpreted as exploratory. No internal validation techniques (e.g., cross-validation, bootstrapping) were applied, as the small pCR sample precludes reliable data splitting.   Figure 1 presents a flowchart outlining patient selection, detailing inclusion and exclusion criteria from initial screening to the final study cohort. Patients were eligible if they had histologically confirmed rectal adenocarcinoma and met criteria for LARC, defined as T3-T4 tumors without lymph node involvement or any T stage with N1-N2 nodal disease, provided no distant metastases were detected on baseline MRI. Additionally, patients required both pre- and post-treatment MRI scans, including DWI and ADC mapping, completion of nCRT and subsequent Total Mesorectal Excision (TME). Nine patients were excluded due to various reasons: Three did not undergo surgical resection or had inadequate histopathological data; three had poor-quality MRI scans or lacked DWI sequences; two exhibited mucinous or signet ring cell histology; and one presented with distant metastases.   Pelvic MRI examinations were scheduled following a fasting period of 5 to 6 hours to reduce intestinal peristalsis. Participants adhered to a clear liquid diet the day before imaging. No bowel preparation, anti-peristaltic agents, or rectal distension methods were employed.     Figure 1: Flowchart of the Study Representing the Criteria and Number of Patients from Initial Retrieval to the Final Study Cohort   Rectal MRI scans were acquired on a 1.5 Tesla Magnetom Era system (Siemens Medical Solutions) using a phased-array body coil, with patients positioned supine. Baseline staging MRI was performed after histopathological diagnosis and re-staging MRI was conducted upon completion of nCRT, employing identical imaging protocols for both assessments. The TME was carried out 6 to 8 weeks following nCRT, followed by histopathological evaluation. MRI protocols included high-resolution T2-weighted imaging in sagittal, axial and coronal planes, alongside diffusion-weighted imaging and ADC mapping.   The MRI protocol comprised oblique sagittal turbo spin echo (TSE) T2-weighted images (TR/TE: 5550/109 ms; field of view [FOV]: 200×100 mm; matrix: 0.6×0.6 mm; slice thickness: 3 mm), oblique axial TSE T2WI (perpendicular to the tumor’s longitudinal axis; TR/TE: 5550/107 ms; matrix: 0.6×0.6 mm; FOV: 200×87 mm; slice thickness: 3 mm), coronal TSE T2WI (TR/TE: 5550/109 ms; matrix: 0.6×0.6 mm; FOV: 200×100 mm; slice thickness: 3 mm) and axial DWI (TR/TE: 8500/68 ms; matrix: 0.99×0.99 mm; FOV: 380×75 mm; slice thickness: 3 mm) with b-values of 0, 50, 400 and 800 s/mm². The ADC maps were generated automatically from DWI data.   Two radiologists, each with 10 to 15 years of experience in abdominal MRI, reviewed the pre- and post-nCRT images in consensus. Tumor staging at baseline was based on high-resolution T2-weighted images, interpreted according to a standardized reporting template [14].   Restaging MRI assessed response and tumor regression using mrTRG (modelled after Mandard histopathologic classification) [15,16] was as follows: mrTRG 1-low signal fibrosis only, no tumor; mrTRG 2-more than 75% fibrosis, minimal tumor; mrTRG 3-50% tumor/fibrosis; mrTRG 4-less than 25% fibrosis, predominant tumor; and mrTRG 5-no fibrosis/regression.   Radiologists delineated Regions of Interest (ROIs) using T2WI and ADC, targeting the largest suspicious tumor area while excluding the lumen, submucosal edema and necrosis. When no residual tumor was apparent on post-nCRT T2WI, ROIs were drawn at the site of the largest initial lesion based on pre-nCRT T2WI.   Pre- and post-nCRT images were transferred in DICOM format to radiomic software (ADC viewer in SYNAPSE, 3D CAD V5.6, FujiFilm). Tumor ROIs were manually delineated on axial ADC maps; the software automatically calculated first-order histogram parameters: mean, minimum, maximum ADC, Standard Deviation (SD), skewness, kurtosis and entropy.   Delta features (Δ) were calculated as post-treatment minus pre-treatment values (Δ Mean ADC, ΔSD, Δ Skewness, Δ Kurtosis, Δ Entropy).   Tumor histopathological response was assessed using the Mandard TRG system, which categorizes response into five levels based on fibrosis and residual tumor presence: TRG1 corresponds to complete regression (fibrosis without tumor), while TRG5 indicates no regression [17]. All assessments were performed by experienced gastrointestinal pathologists. TRG1 cases were defined as pCR, with TRG2-5 classified as non-pCR.   This study intentionally limited analysis to first-order histogram features (mean, minimum, maximum, standard deviation, skewness, kurtosis, entropy). These features describe the distribution of voxel intensities within the ROI without considering spatial relationships between voxels. While second-order (texture matrices such as GLCM, GLRLM) and higher-order (wavelet-transformed) features may capture additional information regarding intratumoral heterogeneity, they were not evaluated in this exploratory study for two reasons. First, the small sample size (n = 34, pCR = 9) precludes stable estimation of higher-order feature models due to the increased risk of overfitting with a larger feature space. Second, first-order features offer direct biological interpretability (e.g., skewness reflecting diffusion histogram symmetry) that facilitates clinical translation.   No internal validation techniques (e.g., k-fold cross-validation, leave-one-out cross-validation, or bootstrap resampling) were performed due to the modest sample size (n = 34) and the limited number of pCR events (n = 9). Data splitting into training and test sets was not feasible, as this would have resulted in an inadequately small test set with insufficient events for meaningful performance evaluation.   Interobserver agreement for ROI placement and texture parameter extraction was not formally assessed in this study. All ROIs were delineated by consensus of two radiologists (with 10-15 years of abdominal MRI experience) following a standardized protocol. While consensus reading reduces systematic variability, the absence of quantitative reproducibility metrics (e.g., intraclass correlation coefficients) is acknowledged as a limitation, particularly for post-treatment scans where tumor boundaries may be indistinct due to fibrosis and edema.   Continuous variables were reported as mean±Standard Deviation (SD) or median with Interquartile Range (IQR), depending on data distribution. Comparisons between groups employed independent t-tests or Mann-Whitney U tests for continuous data and χ² or Fisher’s exact tests for categorical variables. Associations between ADC texture features and pCR were examined by comparing Δ values across groups. Receiver Operating Characteristic (ROC) analyses were performed to assess three predictive models: Model 1 (MRI TRG alone), Model 2 (TRG combined with mean ADC) and Model 3 (comprehensive ADC radiomics incorporating first-order texture features).   The AU (Area Under) ROC, sensitivity and specificity were calculated for each model. A p-value <0.05 was considered statistically significant. Data analyses were performed using SPSS v25 (IBM Corp., Armonk, NY).   This study is reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement for cohort studies.   This study was approved by the Research Ethics Committee of the College of Medicine, Hawler Medical University; Approval No. 25, dated 14 November 2024. Written informed consent was obtained from all participants before inclusion in the study and confidentiality was maintained.

A total of 34 patients with LARC were included in the final data analysis. The mean age was 54.82±12.69 years and 17 patients (50%) were male. Tumors were in the upper rectum in 8 cases (23.5%), mid rectum in 5 cases (17.6%) and lower rectum in 16 cases (47.1%). The median tumor length before treatment was 4.00 cm (IQR, 3.0). Most patients had moderately differentiated adenocarcinoma (88.2%).   Patients achieving pathological complete response (pCR, n = 9) demonstrated greater increases in mean ADC compared to non-pCR patients (0.288±0.116 Vs 0.156±0.178, p = 0.047). Texture analysis revealed that changes in skewness (ΔSkewness: -1.18 Vs -0.41, p = 0.020) and kurtosis (ΔKurtosis: -1.02 Vs -0.60, p = 0.020) were significantly different between groups. Differences in entropy and standard deviation of ADC were not statistically significant (Table 1).   Table 1: Comparison of ADC Texture Features between Complete and Non-Complete Responders Feature (Δ = post–pre) Non-pCR (n = 25) pCR (n = 9) Test p-value Δ Mean ADC (mean±SD) ×10⁻³ mm²/s 0.156±0.178 0.288±0.116 t-test 0.047* Δ SD of ADC (mean±SD) ×10⁻³ mm²/s 0.025±0.155 0.022±0.130 t-test 0.959 Δ Skewness (median [IQR]) –0.41 (–1.19, 0.22) –1.18 (–2.10, –0.92) Mann–Whitney 0.020* Δ Kurtosis (median [IQR]) –0.60 (–0.88, 0.14) –1.02 (–1.47, –0.75) Mann–Whitney 0.020* Δ Entropy (median [IQR]) 0.25 (0.00, 0.56) 0.62 (0.21, 1.09) Mann–Whitney 0.178 Note: ADC, Apparent Diffusion Coefficient, SD, Standard Deviation, IQR, Interquartile Range, pCR, Pathological Complete Response, Δ, change (post-treatment minus pre-treatment). *p<0.05   Figure 2 shows the ROC curve, demonstrating excellent discriminative ability for all three predictive models. The MRI TRG alone model (blue) shows good performance, while the combination of MRI TRG with mean ADC (green) further improves specificity without compromising sensitivity. The full ADC radiomics model (yellow) achieves near-perfect separation, corresponding to an AUC of 1.000, indicating its superior predictive accuracy in distinguishing pCR from non-response.     Figure 2: Receiver Operating Characteristic (ROC) Curves for the Three Predictive Models. PROB1 (blue): MRI TRG Alone. PROB2 (green): MRI TRG Combined with mean ADC. PROB3 (yellow): Full ADC Radiomics Model (First-Order Texture Features). All Models Predict Pathological Complete Response (pCR). Abbreviations: AUC, Area Under the Curve, MRI, Magnetic Resonance Imaging, TRG, Tumor Regression Grade, ADC, Apparent Diffusion Coefficient   ROC analysis demonstrated high diagnostic accuracy of MRI-based models for predicting pCR. MRI TRG alone yielded an AUC of 0.973 (95% CI: 0.923-0.998) with 100% sensitivity (95% CI: 66-100%) and 92% specificity (95% CI: 74-99%). Adding mean ADC values improved the AUC to 0.993 (95% CI: 0.971-1.000) with 100% sensitivity (95% CI: 66-100%) and 96% specificity (95% CI: 80-100%). The full ADC radiomics model achieved an AUC of 1.000 (95% CI: 1.000-1.000) with 100% sensitivity (95% CI: 66-100%) and 100% specificity (95% CI: 86-100%) as presented in Table 2. However, given the limited number of pCR events (n = 9), these perfect metrics likely reflect model overfitting rather than true perfect discrimination. No internal or external validation was performed; therefore, these results are considered hypothesis-generating and require confirmation in larger independent cohorts.   Table 2: Diagnostic Performance of Models for Predicting Pathological Complete Response (pCR) Model Predictors AUC (CI) Sensitivity (%) (CI) Specificity (%) (CI) Comment 1 MRI TRG only 0.973 (0.923–0.998) 100 (66-100) 92 (74-99) Strong baseline model 2 MRI TRG + Mean ADC 0.993 (0.971-1.000) 100 (66-100) 96 (80-100) Improved accuracy 3 Full ADC radiomics model 1.000 (1.000-1.000) 100 (66-100) 100 (86-100) Perfect discrimination AUC, Area Under the Curve, CI, Confidence Interval, MRI, Magnetic Resonance Imaging, TRG, Tumor Regression Grade, ADC, Apparent Diffusion Coefficient   Figure 3 illustrates the difference in the change in Mean ∆ADC between patients who achieved pCR and those who did not (non-pCR). The pCR cohort exhibited a higher median ∆ADC. This observed difference was confirmed with an independent samples t-test. After confirming the assumption of homogeneity of variance via Levene's test (p = 0.297), a statistically significant difference in mean ∆ADC was found, t(32) = -2.064, p = 0.047. The magnitude of the increase in ADC was significantly greater in the pCR group (M = 0.288, SD = 0.116) compared to the non-pCR group (M = 0.156, SD = 0.178).     Figure 3: Boxplot Comparing Δ Mean ADC (change in apparent diffusion coefficient, post-treatment minus pre-treatment) between Patients who Achieved Pathological Complete Response (pCR, n = 9) and Those who did not (non-pCR, n = 25). The Horizontal Line Within Each Box Represents the Median; the Box Spans the Interquartile Range (IQR); Whiskers Extend to the Minimum and Maximum Values. The pCR Group Showed a Significantly Greater Increase in Mean ADC (p = 0.047). Abbreviation: ADC, Apparent Diffusion Coefficient   Table 3 compares ADC values and texture parameters between patients achieving pCR and non-pCR. Both baseline and post-treatment mean ADC were significantly higher in the pCR group, with a greater increase in ΔADC (p<0.001 and p = 0.020). Texture analysis revealed significantly lower post-treatment skewness in pCR patients (p = 0.007), suggesting more symmetrical diffusion histograms following therapy, while kurtosis and entropy remained comparable between groups.   Table 3: Comparison of ADC Values and Texture Parameters between pCR and non-pCR. ADC Parameter Response Group N Mean (±SD) or Mean Rank p-value Mean ADC (×10⁻³ mm²/s) pCR 9 1.094±0.063 <0.001 non-pCR 25 0.958±0.134 Standard Deviation (pre) pCR 9 0.350±0.134 0.852 non-pCR 25 0.340±0.117 Mean ADC (post) ×10⁻³ mm²/s pCR 9 1.382±0.072 <0.001 non-pCR 25 1.114±0.191 Standard Deviation (post) pCR 9 0.372±0.119 0.888 non-pCR 25 0.366±0.115 Δ Mean ADC (post – pre) ×10⁻³ mm²/s pCR 9 0.288±0.116 0.020 non-pCR 25 0.156±0.178 Skewness (pre) pCR 9 22.33 (mean rank) 0.089 non-pCR 25 15.76 (mean rank) Kurtosis (pre) pCR 9 21.89 (mean rank) 0.123 non-pCR 25 15.92 (mean rank) Entropy (pre) pCR 9 13.44 (mean rank) 0.154 non-pCR 25 18.96 (mean rank) Skewness (post) pCR 9 9.89 (mean rank) 0.007 non-pCR 25 20.24 (mean rank) Kurtosis (post) pCR 9 16.22 (mean rank) 0.653 non-pCR 25 17.96 (mean rank) Entropy (post) pCR 9 16.78 (mean rank) 0.800 non-pCR 25 17.76 (mean rank) *Notes: ADC = Apparent Diffusion Coefficient; pCR = Pathological Complete Response, non-pCR = non‑Pathological Complete Response. Normally distributed variables are presented as mean ± standard deviation; non-normally distributed texture parameters are summarized using mean ranks from Mann–Whitney U tests. Significant p‑values (p<0.05) are bolded   Change in Mean ADC demonstrated the highest diagnostic performance, with an area under the ROC curve of 0.776 (95% CI: 0.605-0.947), indicating good discrimination between pCR and non-pCR (Figure 4).     Figure 4: Receiver Operating Characteristic (ROC) Curves for Delta (Δ) ADC Parameters (Post-Treatment minus pre-treatment) in Predicting Pathological Complete Response (pCR). The Δ Mean ADC Curve (Solid Line) Demonstrates the Highest Discriminative Ability, with an Area under the Curve (AUC) of 0.776 (95% CI: 0.605–0.947). Δ Entropy, Δ Skewness and Δ Kurtosis Showed Poor to Fair Discrimination. The Diagonal Dashed Line Represents the Reference Line of no Discrimination (AUC = 0.5). Abbreviation: ADC, Apparent Diffusion Coefficient. Δ Mean ADC is (×10⁻³ mm²/s)   A threshold value of ΔMean ADC ≥ 0.215×10⁻³ mm²/s yielded a sensitivity of 67% (95% CI: 30-93%) and specificity of 72% (95% CI: 51-88%) for predicting pCR.   Changes in texture features, including ΔSkewness, ΔKurtosis and ΔEntropy, showed lower discriminative abilities (AUC = 0.236, 0.236 and 0.653, respectively) (Table 4).   Table 4: ROC Analysis of Δ ADC Parameters for Predicting pCR Parameter AUC (95% CI) Optimal Cut-off (Δ) Sensitivity (%) (CI) Specificity (%) (CI) Interpretation ΔMean ADC 0.776 (0.605–0.947) ≥0.215 67 (30-93) 72 (51-88) Good discrimination ΔSkewness 0.236 (0.062–0.410) — — — Poor discrimination ΔKurtosis 0.236 (0.062–0.410) — — — Poor discrimination ΔEntropy 0.653 (0.463–0.843) — — — Fair discrimination Abbreviations: ADC, Apparent Diffusion Coefficient, AUC, Area Under the Curve, CI, Confidence Interval, Δ, Change (post-treatment minus pre-treatment)

Our study demonstrates that DWI MRI-derived ADC parameters, texture features and tumor morphologic changes predict pCR following neoadjuvant chemoradiotherapy in LARC, serving as noninvasive biomarkers for treatment response assessment and personalized treatment strategies and surgical planning.   Pre-treatment mean ADC and ΔADC% were potential predictors of tumor response in LARC patients undergoing nCRT, with ∆ADC showing excellent performance in distinguishing treatment response and assisting oncologists in optimizing treatment plans [18].   Patients achieving pCR exhibited significantly higher baseline ADC values and greater post-treatment increases compared to non-pCR patients, consistent with evidence that post-treatment ADC elevation reflects reduced tumor cellularity and treatment-induced necrosis [8,19]. Patients achieving pCR exhibited significantly higher baseline ADC values and greater post-treatment increases compared to non-responders, supporting the biological rationale that tumors with higher pretreatment diffusivity and greater ΔADC are more responsive to chemoradiotherapy [20-22]. Wada et al. [23] further demonstrate that ADC changes during nCRT predicted survival outcomes in LARC.   These results strengthen the evidence that ADC is a promising imaging biomarker for treatment monitoring, as it provides functional information on the tumor microenvironment and cellular integrity. Incorporating ΔADC into multiparametric models-alongside volumetric, radiomic and clinical features-could improve treatment response prediction and guide personalized management strategies.   Texture analysis revealed that reductions in skewness and kurtosis were strongly correlated with pCR, while entropy and standard deviation showed no correlation. Histogram-based features reflecting distributional asymmetry and peakedness appear more sensitive to microstructural changes than global heterogeneity measures, consistent with recent studies [8,22].   Pretreatment skewness, kurtosis and entropy showed no significant differences between pCR and non-pCR groups, indicating limited baseline prognostic value. This suggests that baseline tumor heterogeneity on ADC maps may not reliably distinguish responders from non-responders [8,19].   However, post-treatment skewness significantly decreased in the pCR group, suggesting more symmetrical ADC histograms and homogeneous diffusion resulting from necrosis, fibrosis and reduced viable tumor burden [8,24]. Post-treatment kurtosis and entropy showed no significant differences, suggesting lower sensitivity to therapy-induced changes, with recent radiomic studies confirming that skewness is a more sensitive marker of response than kurtosis or entropy [19,24,25].   The limited significance of entropy and standard deviation highlights the necessity of careful feature selection and validation, as not all radiomic features offer equal predictive value [19,26].   Pre- and post-therapy ADC standard deviation did not show significant differences between groups, suggesting intra-tumoral heterogeneity may be less sensitive than mean ADC or ∆ADC [25,27].   These findings support incorporating ADC texture features-particularly skewness-into predictive models, though multicenter validation is needed to confirm their role in personalized treatment strategies [27,28].   ROC analysis demonstrated significant diagnostic accuracy of MRI-based models in predicting pCR. MRI TRG alone showed strong baseline performance [29], with the added mean ADC improving specificity while preserving sensitivity [22].   Δ Mean ADC provided the strongest predictive performance (AUC of 0.776), with a threshold of Δ Mean ADC ≥ 0.215×10⁻³ mm²/s demonstrating balanced sensitivity and specificity, consistent with recent studies identifying ΔADC as highly reliable for predicting LARC treatment response [8,20].   Higher-order texture features (Δ Skewness, Δ Kurtosis) showed poor discriminative ability, while Δ Entropy showed only sufficient performance, suggesting mean ADC changes reflect the most biologically relevant microstructural alterations following chemoradiotherapy [25,27,28].   The full ADC radiomics model achieved perfect discrimination (AUC 1.000), highlighting the capacity of higher-order texture features to enhance predictive performance beyond traditional imaging parameters [22,25,27]. A 2025 meta-analysis confirmed MRI-based radiomics approaches consistently achieve high diagnostic accuracy, supporting their role in personalized treatment planning and organ-preservation strategies [27].   MRI TRG alone Achieved Excellent Accuracy (AUC = 0.973), improving further with mean ADC addition, while the full ADC radiomics model reached perfect discrimination (AUC = 1.000), consistent with meta-analyses showing MRI-based radiomics models outperform conventional MRI in predicting pCR [24,27].   The clear ∆ADC differentiation between pCR and non-pCR groups underscores diffusion-weighted MRI's utility for early response assessment [22,27].   However, ADC thresholds and reproducibility vary across institutions and imaging protocols, necessitating standardization of acquisition and analysis methods for clinical translation [28].   These results support multiparametric MRI combined with radiomics as an effective noninvasive response evaluation tool in LARC, though multicenter validation and workflow standardization remain essential [25,28].   While this sample is modest for radiomics-based modeling, it is comparable to or larger than several published prospective ADC texture analysis studies in LARC, which have reported sample sizes ranging from 21 to 45 patients [8,19,22]. Nevertheless, the small sample-particularly the limited number of pCR events (n = 9)-increases the risk of model overfitting and the perfect discrimination metrics (AUC = 1.000) should be interpreted as hypothesis-generating rather than definitive.   This study provides novel quantitative imaging evidence for radiomics-based response prediction models with implications for broader clinical adoption.   This single-center study with a relatively small cohort may limit generalizability. Although MRI acquisition and analysis were standardized, minor variations in image quality and ROI delineation could introduce measurement bias. Only first-order histogram-based texture parameters were analyzed; higher-order radiomic or deep-learning features may further improve predictive performance. Histopathological assessment was performed by a single team without inter-observer reproducibility testing. External validation using larger, multicenter datasets is required to confirm the robustness and clinical applicability of the proposed model.   While 95% confidence intervals for diagnostic performance metrics have been provided, readers should note that these intervals were calculated using standard methods (e.g., Wilson score for proportions). Due to the small sample size and the presence of perfect separation (100% sensitivity/specificity) in some models, the reported CIs may be artificially narrow and may not accurately reflect the true precision of estimates. For example, the AUC CI of 1.000-1.000 for the full radiomics model is a mathematical artifact of complete separation rather than an indication of infinite precision. This further underscores the need for validation in larger cohorts.   This study did not include a prospective sample size or power calculation. The final cohort of 34 patients (with only 9 pCR events) is relatively small for radiomics-based predictive modeling. The Events-Per-Variable (EPV) ratio, a key consideration in logistic regression and ROC analysis, was substantially below the recommended minimum of 10-20 events per predictor variable. This EPV constraint increases the risk of overfitting and may explain the perfectly discriminative AUC values observed in the full ADC radiomics model. As such, these performance metrics are likely optimistic and should be validated in independent external cohorts before any clinical implementation. Future multicenter studies with prospective power calculations (targeting ≥100 patients with ≥20-30 pCR events) are needed to confirm the robustness of our findings.   A central limitation of this study is the small number of pCR events (n = 9, 26.5% of the cohort). In radiomics and machine learning-based prediction modeling, the Events-Per-Variable (EPV) ratio is a critical determinant of model stability. With only 9 events, even a parsimonious model containing 2-3 predictor variables exceeds the recommended EPV minimum of 10 and our full ADC radiomics model (which considers multiple first-order features) is substantially underpowered. This EPV constraint almost certainly contributes to the perfectly discriminative AUC (1.000) observed for the full radiomics model. Such perfect separation is a known mathematical artifact in small datasets, where a linear combination of features can, by chance, completely separate two groups. We therefore caution readers against overinterpreting the absolute AUC values. External validation in a large, independent cohort with ≥ 50-100 pCR events is mandatory before any clinical use of this model.   This study did not formally evaluate interobserver agreement for ROI placement or texture feature extraction. Although consensus reading by two experienced radiologists was performed to minimize variability, the lack of quantitative reproducibility metrics (e.g., ICC) limits the ability to assess the robustness of first-order histogram parameters. ROI delineation on ADC maps-especially in post-treatment scans where residual tumor may be subtle and admixed with fibrosis-is inherently subjective. Variability in ROI placement could influence skewness, kurtosis and entropy values.

ADC-based radiomic features-particularly changes in mean ADC and post-treatment skewness-may predict pathological complete response after neoadjuvant chemoradiotherapy in locally advanced rectal cancer, with preliminary evidence suggesting potential superiority over morphologic measures alone. However, due to the small number of pCR events (n = 9) and absence of external validation, these findings should be considered exploratory and hypothesis-generating rather than definitive. Integrating ADC metrics with MRI tumor regression grade achieved near-perfect responder discrimination.   These findings support the quantitative diffusion MRI and radiomics for noninvasive, individualized treatment planning and organ preservation. Larger, multicenter validation studies are needed to standardize protocols for clinical translation.   Ethical Statement The study protocol was approved by the Research Ethics Committee of the College of Medicine, Hawler Medical University (Approval No. 25, dated 14 November 2024). All participants provided written informed consent.   Author Contributions Conceptualization, T.H-F.R. and S.A.R.; methodology, T.H-F.R. and S.A.R.; software, not applicable; validation, T.H-F.R. and S.A.R.; formal analysis, T.H-F.R.; investigation, T.H-F.R.; resources, T.H-F.R.; data curation, T.H-F.R.; writing-original draft preparation, T.H-F.R.; writing-review and editing, T.H-F.R. and S.A.R.; visualization, not applicable; supervision, S.A.R.; project administration, T.H-F.R.; funding acquisition, not applicable. All authors have read and agreed to the published version of the manuscript.   Informed Consent Statement Informed consent was obtained from all subjects involved in the study.   Data Availability Statement The datasets generated and analyzed during the current study are available from the corresponding author upon reasonable request.   Acknowledgement The authors acknowledge the staff and patients of the Erbil Thalassemia Care Center for their support and participation in this study.   Conflict of Interest The authors declare that they have no conflicts of interest.   Financial Disclosure This research received no external funding.   Declaration of Generative AI and AI-Assisted Technologies in the Manuscript Preparation Process During the preparation of this work, the authors used (ChatGPT-4), solely for language polishing, SPSS guidance and formatting assistance. All scientific content, data analysis, clinical interpretations and conclusions are the original work of the authors. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.