Artificial Intelligence Models Using Magnetic Resonance Imaging to Predict Response to Chemoradiotherapy in Rectal Cancer: A Systematic Review

ABSTRACT

Background

Pathological complete response (pCR) following neoadjuvant chemoradiotherapy (nCRT) in locally advanced rectal cancer (LARC) is a key prognostic marker with implications for response‐adapted management. Although magnetic resonance imaging (MRI) is central to response assessment, differentiating residual tumour from treatment‐related changes remains challenging. Artificial intelligence (AI) and machine learning (ML) models applied to MRI show promise in predicting pCR; however, variability in methodology and performance limits clinical translation.

Methods

A search of Embase, Medline, Cochrane and Web of Science was conducted in April 2025 in accordance with Preferred Reporting Items for Reviews and Meta‐Analysis (PRISMA) guidelines. Eligible studies used MRI‐only AI or ML models to predict pCR following chemoradiotherapy in adults with rectal cancer. Screening, full‐text review and data extraction were performed independently by two reviewers. Risk of bias was assessed using Quality Assessment of Diagnostic Accuracy Studies (QUADAS‐2).

Results

Twenty‐two studies comprising 94 predictive models were included. Most studies were retrospective, used T2‐weighted MRI and demonstrated variability in MRI protocols, modelling methods and validation strategies. Only five studies conducted external validation. Median AUC was 0.801, with performance ranging from poor to excellent (AUC 0.49–0.997).

Conclusion

MRI‐based AI models demonstrate moderate discriminative performance for predicting pCR following neoadjuvant therapy in LARC. However, methodological heterogeneity, inconsistent reporting and limited external validation currently hinder generalisability. Greater methodological standardisation and multicentre external validation are required before clinical implementation.

1. Introduction

The standard treatment for locally advanced rectal cancer (LARC) (Stages II–III) is neoadjuvant therapy, either conventional neoadjuvant chemoradiotherapy (nCRT) or total neoadjuvant therapy (TNT), followed by total mesorectal excision (TME) [1]. Following chemoradiotherapy, approximately 15%–30% of patients achieve pCR, which is associated with better oncological outcomes, including improved disease‐free survival and lower recurrence rates [2]. However, confirmation of pCR requires surgery, and TME carries a significant risk of morbidity, including permanent stoma formation and low anterior resection syndrome (LARS).

Contrastingly, patients with a clinical complete response (cCR), defined by the absence of residual tumour on digital rectal exam, endoscopy and magnetic resonance imaging (MRI), may undergo a non‐operative “watch‐and‐wait” strategy and avoid surgery [3]. However, cCR remains an imperfect surrogate for pCR. Some patients with apparent cCR have residual microscopic disease, while others without cCR may still achieve pCR.

This diagnostic uncertainty has driven interest in imaging‐based methods to predict pathological response prior to surgery. Although MRI is central to evaluating post‐treatment response, differentiating residual tumour from post‐treatment fibrosis or inflammation remains challenging. Artificial intelligence (AI) and machine learning (ML) techniques have emerged as promising tools to enhance MRI interpretation and quantitatively predict pCR, supporting treatment stratification and organ‐preserving strategies.

ML is a subset of AI that uses algorithms to learn patterns from complex datasets [4]. Radiomics is a quantitative imaging technique involving the extraction of high‐dimensional features from medical images. These features include tumour shape, intensity, texture and spatial heterogeneity and can be used to train ML models [4, 5] to predict pCR. By capturing quantitative imaging characteristics undiscernible to the human eye, radiomics may reflect underlying tumour biology and improve the prediction of treatment response [5]. However, the clinical readiness of AI‐based prediction models remains uncertain.

This systematic review aims to evaluate MRI‐only AI and ML models for predicting pCR in response to neoadjuvant therapy in LARC, focusing on discriminative performance and methodological quality, while summarising patient characteristics, modelling approaches and key limitations to guide further development and translation to clinical use.

2. Methods

2.1. Search Strategy

The literature search was conducted in April 2025 using Embase, Medline, Cochrane Library and Web of Science. Searches were limited to English‐language studies, with no date restrictions. Search strategies used Boolean operators to combine variations of the key terms ‘rectal cancer’, ‘neoadjuvant therapy’, ‘machine learning’ and ‘pathological complete response’. Full search strategies are provided in Tables S1–S4.

2.2. Inclusion and Exclusion Criteria

This review was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta‐Analysis (PRISMA) guidelines. Eligible studies included adult patients (≥ 18 years) with LARC undergoing nCRT or TNT, with subsequent surgical resection and histopathological assessment of pCR. Studies involving short‐course radiotherapy were excluded. Included studies developed AI/ML models using MRI‐derived imaging features to predict pCR. Models incorporating clinical, biochemical or multimodal imaging variables were excluded to isolate the predictive contribution of MRI alone. The primary outcome was discriminative performance, reported as the area under the receiver operating characteristic curve (AUC). Secondary outcomes included methodological characteristics of studies, including modelling approach, feature engineering, neoadjuvant therapy regimen, validation strategy and risk of bias. Detailed inclusion and exclusion criteria are outlined in Table 1.

TABLE 1.

Inclusion and exclusion criteria.

Inclusion criteria

Exclusion criteria

Original research articles (including prospective or retrospective studies, cohort studies, case–control studies and cross‐sectional studies) Adult patients (≥ 18 years) with histologically confirmed LARC treated with conventional neoadjuvant chemoradiotherapy or total neoadjuvant chemoradiotherapy (TNT) Studies using artificial intelligence (AI) or machine learning (ML) models to predict pathological complete response (pCR) Studies that use MRI data only

Non‐human studies or non‐English publications Case reports, review articles, letters, meeting abstracts, editorials or other non‐original research Studies with < 10 patients Studies assessing only chemotherapy or radiotherapy but not both Studies that define treatment response only as ‘responsive’ vs. ‘non‐responsive’ Studies using only radiomic/texture analysis and reporting statistical correlation, without using AI or ML Studies using other imaging modalities apart from MRI (e.g., CT) Studies combining MRI with clinical variables (e.g., CEA)

2.3. Selection Process

Studies were imported into Covidence, and duplicates were removed. Title/abstract and full‐text review were performed independently by W.L. and H.N., with disagreements resolved by consensus or deferral to the senior author, J.Y.

2.4. Data Extraction

Data extraction was performed independently using a customised Covidence template by W.L. and E.E. Disagreements were resolved by consensus discussion. Data extracted included the categories of study characteristics, nCRT regimens, MRI protocols, feature engineering, modelling, validation methods and diagnostic performance.

3. Results

3.1. Study Selection and Characteristics of Included Studies

Our research identified 1046 studies, with 84 studies included for full‐text review and 22 included in the study cohort. The literature search and review process is outlined in Figure 1.

FIGURE 1.

PRISMA flow diagram.

Table 2 provides a summary of the characteristics of eligible studies published between 2019 and 2025, predominantly originating from China (11 studies). Most studies were retrospective (17 studies). The median number of patients in each paper was 156 (range: 40–912). The median pCR rate was 20.8% (range: 16%–73.3%).

TABLE 2.

Study characteristics.

First author and year	Country	Number of centres	Study design	pCR definition	Total number of patients	Training set	pCR (%) training set	Validation set/test set	pCR (%) validation set	pCR (%) total	Validation method
Antunes (2020) [6]	United States of America	3	Retrospective	AJCC	104	NR	NR	T + IV 60 EV 44	T + V 13 (21.67%) EV 10 (22.73%)	23 (22.1%)	IV—50 runs of threefold CV, EV
Bellini (2022) [7]	Italy	1	Prospective	Dworak	40	40	13 (32.5%)	NR	NR	13 (32.5%)	No validation
Bulens (2020) [8]	Belgium	2	Prospective	AJCC	125	70	12 (17.1%)	EV 55	NR	20 (16%)	IV fivefold CV, EV
Crimi (2024) [9]	Italy	1	Prospective	NR	102	72	20 (27.8%)	IV 30	4 (13.3%)	24 (23.5%)	IV—nested 10‐fold CV with split sample
(Ferrari) 2019 [10]	Italy	1	Retrospective	Dworak	55	28	NR	IV 27	NR	16 (29.1%)	IV—split sample
Horvat (2018) [11]	Brazil	1	Retrospective	AJCC	114	NR	NR	NR	NR	21 (18.4%)	IV—fivefold CV
Jang (2021) [12]	South Korea	1	Retrospective	Dworak	466	NR	NR	IV 113 Test 107	20 (17%)	60 (17%)	IV—split sample
Jayaprakasam (2022) [13]	USA	1	Retrospective	AJCC	236	NR	NR	NR	NR	41 (17.4%)	IV—fivefold CV
Lee (2021) [14]	South Korea	1	Retrospective	NR	912	592	114 (19.3%)	Test 320	78 (24.4%)	192 (21.1%)	IV—split sample
Li (2019) [15]	China	1	Retrospective	AJCC	131	87	18 (20.69%)	IV 44	9 (20.45%)	27 (20.6%)	IV—split sample
Li (2021) [16]	China	1	Retrospective	AJCC	80	NR	NR	NR	NR	15 (18.8%)	IV—leave one out CV
Ma (2024) [17]	China	1	Retrospective	NCCN	120	84	NR	IV 36	NR	38 (31.7%)	IV—split sample 70:30
Miranda (2023) [18]	Brazil	1	Retrospective	AJCC	180	126	NR	IV 54	NR	33 (18%)	IV—10‐fold CV with 70:30 split
Nie (2016) [19]	China	1	Prospective	AJCC	48	36	NR	IV 12	NR	11 (23%)	IV—fourfold CV
Pang (2021) [20]	China	2	Retrospective	NCNN	187	107	36 (33.64%)	IV 46 EV 34	IV 8 (17.4%) EV 6 (17.6%)	50 (73.3%)	IV—fivefold CV, EV
Shi (2019) [21]	China	1	Retrospective	AJCC	51	NR	NR	NR	NR	10 (19.6%)	IV—fourfold CV for ANN, 10‐fold CV for CNN
Shi (2023) [22]	China	3	Retrospective (training) Prospective (testing)	AJCC	224	147	41 (27.9%)	EV 77	16 (20.8%)	57 (25.5%)	IV—fivefold CV, EV
Shin (2022) [23]	Republic of Korea	1	Retrospective	Mandard	898	592	NR	IV 306	NR	189 (21%)	IV—10‐fold CV with chronological split
Wu (2025) [24]	China	3	Retrospective	AJCC/CAP	768	401	127 (31.6%)	EV 367	EV 104 (28.3%)	231 (30%)	IV—10‐fold CV, EV
Xia (2025) [25]	China	1	Retrospective	AJCC	373	264	50 (18.9%)	IV 144	IV 22 (15.3%)	72 (19.3%)	IV—split sample
Zhang (2020) [26]	China	1	Prospective	NCNN	383	290	56 (19.3%)	IV 93	18 (19.4%)	74 (19.3%)	IV
Zhu (2022) [27]	China	1	Retrospective	NR	472	200	42 (21%)	IV 72 Test 200	IV 11 (15.3%) Test 44 (22%)	97 (20.6%)	IV—chronological split

Abbreviations: AJCC, American Joint Committee on Cancer; CAP, College of American Pathologists; CV, cross‐validation; EV, external validation; IV, internal validation; NCCN, National Comprehensive Cancer Network; NR, not reported.

3.2. MRI Protocols and Modelling Algorithms

Model construction parameters are summarised in Table S5.

All studies except one utilised T2‐weighted (T2W) imaging. Studies also employed diffusion‐weighted imaging (DWI), T1‐weighted imaging (T1W), dynamic contrast‐enhanced imaging (DCE) and diffusion kurtosis imaging (DKI), in decreasing frequency. Most studies used a 3 T MRI field strength. MRI acquisition timing varied across studies, with some models using pre‐treatment imaging alone, post‐treatment imaging alone or pre‐ and post‐treatment MRI. For MRI segmentation software, most studies used ITK‐SNAP, then 3D‐Slicer and MIM Maestro.

Fifteen studies used ML algorithms alone, including support vector machines (SVM), random forest (RF), least absolute shrinkage and selection operator (LASSO) regression, k‐nearest neighbour (kNN), decision trees (DT), gradient boosting methods and logistic regression, in order of most to least common. Three studies used Deep Learning (DL) algorithms alone, using convolutional neural networks (CNN), whilst four studies used both ML and DL models.

3.3. Validation Methods

Only five studies conducted external validation. The remaining studies conducted internal validation, with k‐fold cross‐validation being the most common, including one study using nested cross‐validation and one using leave‐one‐out cross‐validation. Many studies utilised a split sample for validation, for example, 70:30 training and validation splits or temporal splits.

3.4. Diagnostic Performance

Across 22 studies, 94 models were identified. The mean and median AUC values were 0.778 and 0.801, respectively, with a range of 0.49–0.997 and an IQR of 0.71–0.85. Using commonly accepted AUC interpretation thresholds [28], 13 models were considered excellent (> 0.9), 34 models good (0.8–0.9), 27 models fair (0.7–0.8) and 20 models poor or unacceptable (< 0.7).

3.5. Quality Assessment

Figure 2 shows the QUADAS‐2 assessments. Index test risk of bias had the greatest concern, given that only one paper had prespecified thresholds [9]. The remaining papers either specified thresholds being post hoc or were unclear regarding this. Concerns regarding patient selection bias were present, as four studies either did not state or had unclear inclusion/exclusion criteria. In addition, two studies could have avoided inappropriate exclusions [11, 22].

FIGURE 2.

QUADAS‐2 risk of bias assessment traffic light plot and bar plot.

Concerns regarding the reference standard were generally low, given that all papers used surgical histopathology, the gold standard for determining pCR. Concerns in flow and timing were common, as seven papers were unclear on the interval between nCRT and TME. Applicability concerns were low as the inclusion criteria were already highly specific.

4. Discussion

This review found a median AUC of 0.801, which, although it is considered ‘good’ by conventional thresholds [28], needs to be assessed in context.

4.1. Strengths and Limitations of Selected Studies

External validation is widely regarded as a prerequisite for demonstrating clinical usefulness and generalisability. Only five studies conducted external validation, while the remainder used internal validation. Regarding internal validation types, train‐validation split samples allow validation to be done on unseen data; however, this reduces the effective sample size and increases the risk of variance and instability [29]. Conversely, most studies used resampling techniques such as cross‐validation to maximise the use of limited datasets; however, these models still yield optimistically biased performance estimates [29]. Given the high‐stakes implications of pCR prediction, robust external validation is required for reliability and patient safety.

Our review identified substantial methodological heterogeneity, including differences in MRI protocols, tumour segmentation approaches, feature engineering methods, ML and DL models and validation strategies. This variability limits comparison and generalisability, making it unclear whether differences in reported performance reflect better model design or pipeline variation.

Inconsistent reporting of post‐treatment imaging timing is another limitation. As tumour regression and radiological appearance continue to evolve over time following nCRT [30], variation in imaging timing may influence radiomic feature extraction and model calibration. Consequently, models may learn timing‐related imaging characteristics rather than true biological predictors of response.

To maintain the consistency of the reference standard, only studies using surgical histopathology to show pCR were included, whilst studies including sustained cCR as a surrogate for pCR were excluded.

Heterogeneity in and inconsistent reporting of neoadjuvant treatment regimens were also observed. Reported treatment regimens included long‐course chemoradiotherapy and TNT, with limited reporting of induction or consolidation sequences. As tumour regression is time and treatment‐dependent [30], different neoadjuvant therapy regimens may result in different biological responses and radiological appearances. Most studies did not evaluate the impact of treatment regimen differences on model performance, suggesting that reported predictive accuracy may be regimen‐specific and not broadly generalisable, limiting comparison of models given heterogeneous treatment cohorts.

4.2. Strengths and Limitations of Methodology

This review focuses on MRI‐based AI models to the exclusion of other imaging modalities and clinical variables. Whilst this improves methodological robustness and comparability, this approach does not reflect real‐world practice, whereby MRI is interpreted alongside physical examination, endoscopy, CT and biochemical markers [31] and in the context of the specific treatment regimen used. While high‐resolution MRI remains the preferred modality for local staging and assessment of tumour relationships with the mesorectal fascia and circumferential resection margin [32], optimal rectal cancer management is reliant on multimodal assessment rather than a single imaging technique.

Furthermore, by excluding clinical variables, this review does not compare the performance of clinical‐only, imaging‐only and combined prediction models. Consequently, the incremental predictive value of MRI‐based AI beyond established clinical predictors of treatment response cannot be determined. Readily available clinical features such as carcinoembryonic antigen (CEA) levels and tumour stage can be easily incorporated into combined models and may substantially influence predictive performance. Direct comparison of clinical, imaging and combined approaches is therefore required to determine whether MRI‐based AI meaningfully improves discrimination and has clinical utility in guiding patient selection for organ preservation strategies.

Significant heterogeneity across studies precluded meta‐analysis. As a result, results were synthesised descriptively, limiting statistical analysis and the strength of evidence compared with meta‐analyses.

ML and DL differ primarily in feature extraction. ML approaches rely on predefined features derived from imaging data, whereas DL approaches use multi‐layered neural networks to automatically learn features directly from imaging data [5]. ML models may perform well in small datasets and offer greater interpretability but are dependent on feature selection and may be sensitive to variation in MRI protocols [33]. In contrast, DL models can capture complex spatial relationships and may be advantageous in distinguishing residual tumour from treatment‐related fibrosis. However, DL approaches require larger datasets and have a higher risk of overfitting in small, single‐centre cohorts [34], which characterises most studies in this review. Consequently, differences in model performance are likely influenced by dataset size, standardisation and validation strategy in addition to algorithm type. The heterogeneity of included studies precluded direct comparison between ML and DL. However, prior reviews have suggested that DL models may demonstrate improved performance [35, 36, 37].

4.3. Comparison With Prior Literature

Prior literature has reported higher AUCs (approximately 0.9–0.91) than our median of 0.801; however, these reviews frequently incorporated clinical features, other imaging modalities, or selectively reported high‐performing models. Jia et al. [35] and Jong et al. [38] demonstrated higher median AUCs whilst incorporating clinical features, whilst Liao et al. [37] and Shen et al. [39] reported higher performance by including clinical features and selection of the validation/test cohorts with ‘superior predictive performance’ or including other imaging modalities, respectively. Subgroup analysis by Jia et al. [35] found that incorporating clinical factors yielded higher diagnostic accuracy than with radiomics models alone. Studies incorporating clinical features or other imaging modalities may achieve higher AUCs and have greater predictive performance. However, inclusion of these factors limits the ability to isolate the predictive contribution of MRI and reduces comparability between MRI‐based modelling pipelines.

Shin et al. [23] demonstrated that AI models surpass radiologists' visual assessment, whilst Zhang et al. [26] demonstrated that AI models can significantly reduce radiologist error rates as an assistive tool. Whilst the predictive performance of imaging alone is modest at 0.801 in this review, it would likely still be a useful assistive tool for radiologists' visual assessment.

4.4. Implications of the Study/Future Directions

This review highlights the need for greater methodological standardisation and transparent reporting to improve reproducibility in MRI‐based radiomics research. Future studies should prioritise standardised radiomics feature extraction and reporting, adhering to established frameworks such as the Image Biomarker Standardisation Initiative (IBSI) and TRIPOD‐AI reporting guidelines [40, 41].

Most studies identified in this review remain at the retrospective model development stage. Future research should prioritise external validation using multicentre cohorts to ensure models are robust and generalisable across different populations and imaging protocols. Subsequent prospective studies are required, including studies to evaluate the clinical utility of these models and their integration into multidisciplinary decision‐making pathways, particularly in guiding patient selection for organ preservation strategies. Future studies should also evaluate whether MRI‐based AI models provide incremental predictive value beyond established clinical predictors.

5. Conclusion

This review demonstrates that MRI‐based AI models achieve moderate discriminative performance for predicting pCR in LARC. However, these findings should be interpreted cautiously, as performance estimates are likely to be optimistically biased due to reliance on small, retrospective cohorts with internal validation strategies. Methodological heterogeneity and inconsistent reporting further limit generalisability. Standardised methodology and robust multicentre external validation are essential for safe clinical translation to position MRI‐based AI models as an assistive component to multimodal assessment of LARC treatment response.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Table S1: Ovid Medline search strategy.

Table S2: Embase search strategy.

Table S3: Cochrane search strategy.

Table S4: Web of Science search strategy.

Table S5: Modelling construction parameters.

Table S6: Model performance.

Table S7: Chemoradiotherapy regimens.

Data Availability Statement

The data that supports the findings of this study are available in the Supporting Information of this article.