Multi‐Sequence Fusion MRI Deep Learning Models for Discrimination of Recurrent Nasopharyngeal Carcinoma and Osteoradionecrosis

ABSTRACT

Objectives

To develop a multi‐sequence fusion model based on MRI images to differentiate between recurrent nasopharyngeal carcinoma (rNPC) and osteoradionecrosis (ORN) in patients with nasopharyngeal carcinoma (NPC) after radiotherapy, and to evaluate its diagnostic performance.

Methods

We retrospectively reviewed 370 patients with pathologically confirmed recurrent nasopharyngeal carcinoma (rNPC) or osteoradionecrosis (ORN). After screening, 346 patients (207 rNPC and 139 ORN) with both axial T1‐weighted contrast‐enhanced imaging (T1C) and T2‐weighted imaging (T2WI) were included and randomly split into a training set (n = 212), validation set (n = 61), and test set (n = 73). A dual‐branch 3D‐ResNet‐50‐based fusion model was developed to integrate multi‐sequence MRI features. Model performance was evaluated using AUC, accuracy, sensitivity, specificity, precision, and F1 score.

Results

On the test set, the proposed 3D‐ResNet‐Fusion model achieved an AUC of 0.86 and an accuracy of 84%, with a sensitivity of 91%, specificity of 72%, precision of 83%, and an F1 score of 0.87. Based on AUC comparisons, the fusion model outperformed comparator models based on EfficientNet (0.77, p = 0.20) and DenseNet (0.70, p = 0.02), as well as single‐sequence 3D‐ResNet models trained on T1C alone (0.71, p = 0.02) or T2WI alone (0.73, p = 0.046). In the human–AI comparison experiment, the proposed AI model achieved a 3% higher ACC than the senior otolaryngologist (p = 0.83), and outperformed the two junior otolaryngologists by 15% (p = 0.05) and 22% (p < 0.01), respectively.

Conclusion

The proposed AI model effectively integrates information from multiple MRI sequences, significantly improving classification performance in distinguishing rNPC from ORN.

Key Points

• This study successfully developed a multi‐sequence fusion model based on MRI images, demonstrating its potential to effectively differentiate between recurrent nasopharyngeal carcinoma and osteoradionecrosis in patients with nasopharyngeal carcinoma who have undergone radiotherapy.

• The proposed AI model is a reliable diagnostic tool, and its accurate assessment of recurrent nasopharyngeal carcinoma and osteoradionecrosis aids in developing tailored clinical strategies for different patients, thereby supporting patient prognosis.

1. Introduction

Nasopharyngeal carcinoma (NPC) is a malignant tumor with a high incidence in East and Southeast Asia. Radiotherapy remains the primary treatment modality for newly diagnosed NPC patients [1, 2]. Despite significant improvements in therapeutic efficacy with the widespread implementation of intensity‐modulated radiotherapy (IMRT) in recent years, approximately 10% of patients still experience recurrence at the primary site or in the local region [3, 4]. Meanwhile, radiotherapy may also induce necrosis, with more severe cases progressing to skull base osteoradionecrosis (ORN) [5, 6]. Some recurrent Nasopharyngeal carcinoma (rNPC) cases involve lesions that may or may not be accompanied by necrosis, forming complex pathological states [7]. These complex lesions are particularly challenging to distinguish from isolated ORN in clinical practice. As a result, patients with isolated ORN may be mistaken for having recurrent nasopharyngeal carcinoma and may therefore receive unnecessary high‐risk salvage treatments, including surgery or re‐irradiation, potentially leading to severe complications such as internal carotid artery rupture and substantially impairing quality of life [8, 9]. On the other hand, patients with recurrent disease who are misdiagnosed as having isolated ORN may miss the optimal window for intervention, leading to disease progression and reduced survival rates [10].

Currently, the primary examination methods for monitoring NPC patients after radiotherapy include nasopharyngoscopy, MRI, PET‐CT, and surgical biopsy [11, 12]. Nasopharyngoscopy may fail to detect submucosal tumors, while PET‐CT, despite its high cost, offers specificity comparable to MRI, making neither method suitable for routine post‐radiotherapy monitoring of NPC [13, 14]. Given MRI's advantages of high soft‐tissue contrast, multi‐planar imaging capability, and relatively manageable cost, it has become the preferred imaging modality for NPC surveillance following radiotherapy [15]. However, distinguishing rNPC at the primary site from ORN on MRI remains a challenge due to their similar imaging characteristics [16]. Furthermore, suspected recurrent lesions are often located deep within the nasopharynx, rendering biopsies highly invasive and associated with substantial surgical risks [17]. Consequently, there is an urgent need for novel tools to accurately differentiate these pathological conditions.

In recent years, deep learning has been widely applied in the field of medical imaging, thanks to its multi‐layer and end‐to‐end architecture [18, 19]. It has demonstrated higher accuracy and efficiency compared to traditional methods in disease diagnosis [20], lesion segmentation [21], tumor detection [22], and prognosis prediction [23]. However, current artificial intelligence (AI) research on MRI primarily focuses on single sequences, missing the value of combining multiple sequences for diagnosis.

To address those limitations, this study aims to develop a deep learning‐based multi‐sequence fusion diagnostic algorithm for distinguishing rNPC from ORN. The reliability of the proposed algorithm is validated through comparisons between different backbones, between single‐ and multimodal approaches, and between the AI model and otolaryngologists.

2. Materials and Methods

2.1. Data sets

This retrospective study was approved by the ethics committee of the Eye & ENT Hospital, Fudan University (2019095‐2), with informed consent waived. This was a single‐center retrospective study for model development and internal validation, and it was reported in accordance with the TRIPOD + AI guidelines where applicable [24]. We retrospectively retrieved records of patients who underwent nasopharyngeal lesion resection surgery at the hospital between February 2017 and February 2025. Patients were included if they met the following criteria: (1) a history of primary NPC that achieved complete remission following radiotherapy; (2) postoperative pathology confirming either NPC or ORN; (3) preoperative MRI scans including axial T1‐weighted contrast‐enhanced imaging (T1C) and T2‐weighted imaging (T2WI) sequences; and (4) images of sufficient quality clearly depicting the lesion. Exclusion criteria were as follows: (1) pathology report indicates unclear findings; (2) missing essential MRI sequences; (3) images with insufficient quality or significant artifacts; (4) presenting with other malignancies or distant metastasis of NPC; and (5) duplicate cases. The screening process was rigorously conducted by an otolaryngologist (Dr. Shi, with 5 years of clinical experience). Figure 1 is the flowchart.

Figure 1.

Flowchart of the data selection process. ORN, osteoradionecrosis; rNPC, recurrent Nasopharyngeal carcinoma.

After screening, a total of 370 patients were included in the study. Their demographic and clinical characteristics, including gender and age, were collected for those in the final cohort. Detailed patient information is summarized in Table 1. All patients underwent preoperative contrast‐enhanced nasopharyngeal MRI at our hospital. They were randomly allocated to one of two MRI scanners: the Magnetom Verio (3.0 T, Siemens Healthcare, Erlangen, Germany) with a 12‐channel head and neck coil, or the Magnetom Prisma (3.0 T, Siemens Healthcare, Erlangen, Germany) with a 64‐channel head and neck coil. Scanning parameters are also detailed in Table 1.

Table 1.

Demographic and imaging information of the three sets.

Items		Train set (n = 212)				Validation set (n = 61)				Test set (n = 73)
	rNPC	ORN	Statistic	p‐value	rNPC	ORN	Statistic	p‐value	rNPC	ORN	Statistic	p‐value
Age (years, Mean ± SD)	51 ± 11	56 ± 10	t = −3.33	< 0.01	50 ± 11	55 ± 10	t = −1.71	0.09	52 ± 13	56 ± 9	t = −1.64	0.11
Sex[n(%)]
Male	98 (78)	64 (74)	${\chi }^2$ = 0.16	0.69	25 (68)	20 (83)	${\chi }^2$ = 1.14	0.29	33 (75)	23 (79)	${\chi }^2$ = 0.02	0.89
Female	28 (22)	22 (26)			12 (32)	4 (17)			11 (25)	6 (21)
T1C
TR (ms)	3.97	3.92	—	—	4.11	3.90	—	—	4.04	3.86	—	—
TE (ms)	1.63	1.62			2.00	1.60			1,71	1.54
Pixel Spacing (mm)	0.67	0.68			0.64	0.68			0.64	0.69
T2WI
TR (ms)	4030	4154	—	—	3902	4258	—	—	3724	4163	—	—
TE (ms)	88	89			87	91			88	91
Pixel Spacing (mm)	0.59	0.59			0.54	0.57			0.59	0.59

Note: —, no data.

Abbreviations: ORN, osteoradionecrosis; rNPC,recurrent Nasopharyngeal carcinoma; T1C, T1‐weighted contrast‐enhanced imaging; T2WI, T2‐weighted imaging.

2.2. Image Preparation

We first exported imaging data from PACS in DICOM format, then identified the T1C and T2WI sequences for each case. These two sequences were converted into NII format files and imported into our image preprocessing pipeline. The first step involved normalizing all MRI images by fixing the signal intensity within the range of 0–4000 arbitrary units and then normalizing it to 0–1, which enhances the model's training efficiency. In the second step, the size and dimensions of the MRI images were adjusted. The target depth, width, and height were set to 40, 100, and 100 voxels, respectively. The images were resized along the z‐axis based on the calculated scaling factor, and their dimensions were expanded to ensure consistent input size and facilitate subsequent feature concatenation required by the model. In the third step, during training, a random rotation angle (−20° to 20°) was applied to each MRI sequence prior to model input. This data augmentation improves the model's generalization ability and reduces the risk of overfitting during training.

2.3. Fusion Model Construction

In this study, we propose a dual‐branch fusion network comprising two parallel feature extraction branches integrated with a multimodal fusion module. Each feature extraction branch employs 3D‐ResNet‐50 as its backbone [25], leveraging residual connections to overcome performance degradation in deep networks. These skip connections facilitate efficient gradient backpropagation to shallow layers, effectively addressing vanishing gradient problems. The 3D convolutional architecture inherently maintains volumetric spatial relationships in MRI data, reducing dependency on manual lesion annotations while enhancing cross‐plane feature learning of lesion characteristics. To optimize performance for our specific diagnostic task, we implemented targeted modifications to the baseline architecture.

As illustrated in Figure 2, the model comprises four key components: (1) an input layer for volumetric data processing, (2) four hierarchical stages containing 16 convolutional operations within residual blocks for multi‐scale feature extraction, (3) a cross‐modal feature fusion module integrating information from both branches, and (4) a classification head with task‐specific output layers.

Figure 2.

The Working Framework of Fusion Model. T1C and T2WI images of recurrent nasopharyngeal carcinoma (rNPC) and osteoradionecrosis (ORN) are separately fed into two 3D‐ResNet‐50 backbone networks. After passing through a Multimodal Fusion Module, the features are merged and then processed by a classification head to output the probabilities of rNPC and ORN.

In the input layer, we adjusted the channel count of the initial 7 × 7 convolution (stride = 2, output channels = 64) from 3 to 1 to process grayscale MRI images, followed by a 3 × 3 max pooling layer (stride = 2).

Each stage consists of multiple Residual Blocks: 3 in stage 1, 4 in stage 2, 6 in stage 3, and 3 in stage 4. Each Residual Block integrates three convolutional operations with a residual connection. Specifically, a 1 × 1 convolution is initially applied to reduce the channel dimensions, thereby improving computational efficiency. This is followed by a 3 × 3 convolution that captures local spatial features, and subsequently, another 1 × 1 convolution is used to expand the channel dimensions, enhancing the model's ability to represent complex features. Finally, the input features of the Residual Block are combined with the output of the three convolutional layers through element‐wise addition, producing the final feature representation.

To integrate multimodal information from two MRI sequences, we designed a multimodal fusion module (as shown in Figure 2). After global average pooling for both modalities, we concatenated the deep features extracted by the network along the same dimension, creating a unified feature tensor. Subsequently, a fully connected layer was employed to facilitate feature‐level interactions between modalities, enabling the network to learn cross‐modal correlations and generate a compact and discriminative representation for downstream classification.

Finally, the unified feature embedding from the multimodal fusion module was passed to the classification head, where a fully connected layer projects the fused features into a task‐specific representation, followed by a sigmoid layer that outputs class probabilities for disease classification.

During training, the learning rate was set to 0.0005, and BCE Loss was chosen as the loss function, with the maximum number of epochs fixed at 400. Model parameters were optimized using the Adam optimizer with a batch size of 8. A fivefold cross‐validation approach was adopted, and the parameters corresponding to the epoch achieving the highest prediction accuracy on the validation set were preserved for further evaluation.

2.4. Evaluation Metrics

During the model performance validation phase, all preprocessing steps and parameter settings were first ensured to be consistent across validation groups. Three types of validation were conducted: comparison of different model architectures and fusion strategies, and a human–AI comparison involving otolaryngologists. rNPC was designated as the positive class, and all comparative experiments were evaluated using a comprehensive set of metrics, including accuracy, area under the receiver operating characteristic (ROC) curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the F1 score.

The first validation focused on evaluating the performance of the backbone network. 3D‐DenseNet‐121 and 3D‐EfficientNet‐B1 were selected as baseline networks for comparison [26, 27], where all components except the feature concatenation and classification head were replaced. A default probability threshold of 0.5 was used for binary classification. In addition, threshold‐based performance analysis was performed to evaluate changes in model performance under different probability thresholds. Decision curve analysis (DCA) was performed to assess the clinical net benefit of the 3D‐ResNet‐Fusion model across different threshold probabilities, using treat‐all and treat‐none strategies as reference standards.

The second validation examined the effectiveness of modality fusion. Using the standard 3D‐ResNet‐50 architecture, we trained and validated the model separately on patient T1C and T2WI image sequences. Statistical comparisons among different AI models were based on AUC. To further evaluate the interpretability of the fusion model, gradient‐weighted class activation mapping (Grad‐CAM) was applied to the trained 3D‐ResNet‐Fusion model. For each MRI sequence branch, feature maps were extracted from the final residual block of the last convolutional stage, and gradients of the predicted class score with respect to these feature maps were used to generate class‐discriminative heatmaps. The resulting heatmaps were overlaid on the original T1C and T2WI images to visualize image regions that contributed to model predictions.

Finally, we compared the classification performance of the AI model with that of one otolaryngologist with over 10 years of experience and two otolaryngologists with less than 5 years of experience. All otolaryngologists were blinded to the histopathological results of each case, and the cases were randomly shuffled. Based on the T1C and T2WI images of each case, the otolaryngologists determined whether the lesion represented rNPC or ORN. Statistical comparisons among different AI models were based on AUC. Statistical comparisons between the AI model and otolaryngologists were based on accuracy.

3. Results

3.1. Patient Characteristics

A total of 370 patients were included in this study. 24 patients were excluded due to unclear pathological reports (n = 4), imaging artifacts (n = 3), missing essential MRI sequences (n = 8), presenting with other malignancies or distant metastasis of NPC (n = 7), or duplicate cases (n = 2), as shown in Figure 1. As a result, 346 patients were included in the final cohort, consisting of 207 with rNPC and 139 with ORN. The cohort was partitioned into training (n = 212), validation (n = 61), and test (n = 73) subsets in a class‐balanced manner so that the proportions of ORN and rNPC remained comparable across the three groups. The mean age of the patients was (53 ± 11) years, and 76% of them were male. Among the analyzed variables, except for the age difference between the two groups in the training set (p < 0.01), all other variables did not show statistical significance (p > 0.05). Detailed demographic characteristics are summarized in Table 1.

3.2. AI Model Performance

As illustrated in Table 2 and Figure 3, the 3D‐Resnet‐Fusion model demonstrated strong performance on both the validation and test sets. Specifically, the validation set achieved an ACC of 89% (95% CI: 80%, 96%) and an AUC of 0.87 (95% CI: 0.75, 0.97). For the test set, the ACC was 84% (95% CI: 75%, 92%), and the AUC was 0.86 (95% CI: 0.76, 0.95). Both the validation and test sets demonstrated high sensitivity, with values of 91% (95% CI: 79%, 100%) and 91% (95% CI: 82%, 98%), respectively. Additionally, we evaluated different backbone networks, showing that 3D‐EfficientNet and 3D‐DenseNet achieved test set AUC of 0.78 (95% CI: 0.65, 0.89) and 0.80 (95% CI: 0.70, 0.90), respectively, with p‐values of 0.20 and 0.02 when compared to 3D‐Resnet. The classification confusion matrices, including true positives, false positives, true negatives, and false negatives for all AI models on the test set, are presented in Figure 4. As shown in Figure 5, threshold‐based analysis showed that the sensitivity and specificity of the 3D‐ResNet‐Fusion model varied across different probability thresholds. Decision curve analysis showed that the 3D‐ResNet‐Fusion model had a higher net benefit than the treat‐all and treat‐none strategies across most threshold probabilities, as shown in Figure 6.

Table 2.

Performance of models in the validation set and test set.

Items	AUC	Statistic	p‐value	Accuracy (%)	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	F1 score
Validation set
3D‐Resnet‐Fusion	0.87 [0.75, 0.97]	—	—	89 [80, 96]	91 [79, 100]	86 [69, 100]	91 [79, 100]	86 [70, 100]	0.91
3D‐Efficientnet‐Fusion	0.78 [0.64, 0.91]	—	—	80 [69, 91]	84 [70, 97]	73 [53, 90]	82 [67, 94]	76 [57, 94]	0.83
3D‐Densenet‐Fusion	0.84 [0.73, 0.95]	—	—	83 [72, 93]	88 [75, 97]	77 [56, 94]	85 [71, 96]	81 [63, 96]	0.86
3D‐Resnet‐T1C	0.87 [0.74, 0.98]	—	—	87 [78, 96]	88 [74, 97]	86 [69, 100]	90 [79, 100]	83 [63, 96]	0.89
3D‐Resnet‐T2WI	0.88 [0.79. 0.98]	—	—	87 [78, 94]	91 [79, 100]	82 [65, 96]	88 [76, 97]	86 [68, 100]	0.89
Test set
3D‐Resnet‐Fusion	0.86 [0.76, 0.95]	—	—	84 [75, 92]	91 [82, 98]	72 [55, 89]	83 [73, 94]	84 [68, 96]	0.87
3D‐Efficientnet‐Fusion	0.78 [0.65, 0.89]	−1.29	0.20	77 [67, 86]	89 [67, 98]	59 [40, 77]	77 [65, 88]	77 [58, 94]	0.82
3D‐Densenet‐Fusion	0.80 [0.70, 0.90]	−2.36	0.02	70 [59, 80]	75 [61, 87]	62 [43, 79]	75 [62, 88]	62 [44, 79]	0.75
3D‐Resnet‐T1C	0.80 [0.69, 0.90]	−2.32	0.02	71 [60, 81]	77 [64, 89]	62 [44, 79]	76 [63, 88]	64 [46, 81]	0.76
3D‐Resnet‐T2WI	0.78 [0.67, 0.88]	−2.00	0.046	73 [63, 82]	77 [64, 89]	66 [48, 83]	77 [64, 90]	66 [47, 82]	0.77

Note: —, no data; Data within the brackets represent 95% confidence intervals. p‐values indicate the statistical significance of AUC differences between the 3D‐ResNet‐Fusion model and each of the comparator models.

Abbreviations: AUC, area under the ROC curve; F1 score, Harmonic mean of precision and recall; NPV, negative predictive value; PPV, positive predictive value; T1C, T1‐weighted contrast‐enhanced imaging; T2WI, T2‐weighted imaging.

Figure 3.

The ROC curve of five models on the test sets.

Figure 4.

The confusion matrix of the AI model on the test set. Recurrent nasopharyngeal carcinoma was defined as the positive class.

Figure 5.

Threshold‐based performance analysis of the 3D‐ResNet‐Fusion model. The curves show changes in accuracy, sensitivity, specificity, and F1 score across different probability thresholds on the test set. The dashed vertical line indicates the default threshold of 0.5.

Figure 6.

Decision curve analysis of the 3D‐ResNet‐Fusion model. Decision curve analysis was performed on the test set to compare the net benefit of the 3D‐ResNet‐Fusion model with treat‐all and treat‐none strategies across different threshold probabilities. rNPC was defined as the positive class.

3.3. Fusion Model Performance Compared With Single‐Modality Networks

Additionally, we evaluated the performance of the multimodal fusion module. On the validation set, the single‐modality networks demonstrated performance comparable to the fused‐modality network, with T1C achieving an ACC of 87% (95% CI: 78%, 96%) and T2WI achieving an ACC of 87% (95% CI: 78%, 94%). However, on the test set, the performance of the single‐modality networks declined markedly. T1 achieved an ACC of 71% (95% CI: 60%, 81%) and an AUC of 0.80 (95% CI: 0.69, 0.90), while T2 achieved an ACC of 73% (95% CI: 63%, 82%) and an AUC of 0.78 (95% CI: 0.67, 0.88). The p‐values for AUC comparisons between the fusion model and the T1C and T2WI models were 0.020 and 0.046, respectively. Grad‐CAM heatmaps showed that the 3D‐ResNet‐Fusion model mainly focused on lesion‐related regions in the nasopharynx and skull base area on T1C and T2WI images, as shown in Figure 7.

Figure 7.

Grad‐CAM heatmaps of a representative case generated by the 3D‐ResNet‐Fusion model. (A) T1C image and corresponding Grad‐CAM heatmap and (B) T2WI image and corresponding Grad‐CAM heatmap. In each panel, the original MRI image is shown on the left and the heatmap overlay is shown on the right. The color bar represents normalized Grad‐CAM intensity, with higher values indicating greater contribution to the model prediction.

3.4. AI Model Performance Compared With Otolaryngologists

In the human–AI comparison, as shown in Table 3 and Figure 8, the 3D‐Resnet‐Fusion model slightly outperforms the senior otolaryngologist (DOC1: 81%, 95% CI: 73%, 89%, p = 0.83), and significantly outperforming the two junior otolaryngologists (DOC2: 69%, 95% CI: 58%, 80%, p = 0.05; DOC3: 62%, 95% CI: 50%, 73%, p < 0.01). In terms of specificity, the 3D‐Resnet‐Fusion model achieved 72%, compared to 55% for DOC1, 48% for DOC2, and 41% for DOC3.

Table 3.

Comparison of AI model and otolaryngologists' performance.

Items	Accuracy (%)	Statistic	p‐value	Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	F1 score
3D‐Resnet‐Fusion	84 [75, 92]	—	—	91 [82, 98]	72 [55, 89]	83 [73, 94]	84 [68, 96]	0.87
DOC1	81 [73, 89]	0.05	0.83	98 [92, 100]	55 [36, 73]	77 [66, 88]	94 [80, 100]	0.86
DOC2	69 [58, 80]	3.70	0.05	82 [69, 93]	48 [28, 66]	71 [59, 83]	64 [42, 84]	0.76
DOC3	62 [50, 73]	8.65	< 0.01	75 [63, 87]	41 [24, 60]	66 [54, 80]	52 [32, 72]	0.70

Note: —, no data; DOC1 denotes the senior otolaryngologist; DOC2 and DOC3 denote junior otolaryngologists. p‐values indicate the statistical significance of accuracy differences between the 3D‐Resnet‐Fusion model and each otolaryngologist.

Abbreviations: NPV, negative predictive value; PPV, positive predictive value.

Figure 8.

ROC curves of AI system versus otolaryngologists for rNPC–ORN differentiation.

4. Discussion

In this study, we developed a fusion‐based deep learning framework to differentiate rNPC from ORN using multi‐sequence MRI. The model achieved an ACC of 84% and an AUC of 0.86 on the test set, significantly outperforming two junior otolaryngologists and showing a slight advantage over a senior otolaryngologist, demonstrating its effectiveness in diagnosing rNPC and ORN.

In clinical practice, isolated ORN without complications is typically managed with anti‐inflammatory treatment and nasal irrigation, along with periodic monitoring, whereas patients with suspected rNPC are usually treated with surgery or re‐irradiation [28]. Studies have shown that the 2‐year overall survival (OS) rates for patients undergoing surgery are 83.9% for rT1–2 and 35.6% for rT3–4, indicating that early‐stage rNPC is more likely to benefit from intervention and achieve longer survival [29]. As a result, clinicians may tend to diagnose suspicious lesions as rNPC—a defensive strategy driven by ethical concerns to minimize the risk of missing malignant recurrence, even at the expense of increased harm to patients. This tendency is reflected in the relatively low specificity of otolaryngologists, with DOC1, DOC2, and DOC3 achieving specificity rates of 55%, 48%, and 41%, respectively. Furthermore, due to the geographically uneven distribution of NPC prevalence, diagnostic imaging capabilities for detecting recurrence and its complications are limited in certain regions. These limitations may contribute to incorrect treatment strategies and potentially adverse outcomes for patients. In this context, the proposed AI model may help provide more accurate and personalized treatment strategies for patients, offering valuable support in complex diagnostic scenarios.

We conducted a search on PubMed for articles published in the past decade related to AI‐assisted diagnosis of NPC and related diseases using MRI. Few studies leveraged multiple sequences MRI obtained from a single examination of the patient. Among them, Wong et al. [30] constructed a three‐dimensional Residual Attention Network using T2‐weighted MRI of patients to differentiate between NPC and benign hyperplasia in the nasopharynx, achieving promising performance (ACC = 91.5%). Hua et al. [31] collected post‐radiotherapy T1 contrast‐enhanced MRI images of NPC patients to build a convolutional neural network for predicting residual tumor risk associated with different treatment modalities, also yielding favorable outcomes (ACC = 84.1%). In clinical practice, the T1WI sequence is typically used to visualize anatomical structures, while the T2WI sequence is better at highlighting lesions and pathological features. However, traditional neural networks often focus on a single sequence and fail to integrate complementary information across multiple sequences, resulting in a loss of diagnostic details in patient imaging data. Moreover, in single‐modality networks, the accuracy of the model may be affected by noise or artifacts, leading to image quality degradation and reduced robustness.

To address these challenges, we developed a multimodal fusion module at the tail of our model, designed to comprehensively integrate imaging information from both sequences. Compared with traditional single‐modality neural networks, our new model incorporates more imaging information and demonstrates superior classification ability. On the test set, the accuracy of the single‐modality networks (T1C: 71%, T2WI: 73%) was improved to 84% by our model. In addition, Grad‐CAM heatmaps showed that the fusion model focused on lesion‐related regions in both T1C and T2WI images, further indicating that the model extracted relevant information from both modalities. This finding suggests that future studies using multi‐sequence MRI could explore multi‐modal networks to enhance the performance of classification tasks.

Although our study demonstrated good performance on the test set, it has certain limitations. First, the model was not trained or validated on external center data, which may limit its generalizability. To address this, we plan to enhance generalizability by including data from additional centers in future studies. Second, our model focuses exclusively on MRI imaging data, whereas in real‐world clinical practice, physicians also consider clinical information (e.g., age, gender, and medical history) when evaluating patient conditions. Future work will explore integrating clinical information into the classification network. Finally, although the proposed model was compared with otolaryngologists, comparison with radiologists was not included in the current study. Future studies should include radiologists with different levels of experience to further evaluate the clinical value of the model in imaging‐based diagnosis.

In summary, this study developed a fusion‐based convolutional neural network model using multi‐sequence MRI imaging to distinguish between rNPC and ORN, achieving encouraging performance. This marks the first attempt to differentiate these two pathological conditions and shows promising potential for clinical applications. Future research should delve deeper, leveraging larger data sets, diverse data modalities, and multi‐center collaboration to better support prognostic decisions for NPC patients.

Author Contributions

Wen‐Tao Gong: conceptualization, data curation, formal analysis, methodology, software, validation, visualization, writing – original draft, writing – review and editing. Xue‐Li Liu: formal analysis, methodology, software, validation, writing – review and editing. Xiao Chen: data curation, investigation, validation. Zhen Li: data curation, investigation. Ming‐Yan Qiu: data curation, investigation. Shang Shi: data curation, investigation. Xuan‐Yu Zhao: data curation, investigation. Li Wang: data curation, funding acquisition, investigation, resources. Quan Liu: conceptualization, project administration, supervision, writing – review and editing. Xi‐Cai Sun: investigation, resources. Yu‐Xuan Shi: conceptualization, funding acquisition, methodology, project administration, supervision, writing – review and editing. Hong‐Meng Yu: conceptualization, funding acquisition, project administration, resources, supervision, writing – review and editing.

Ethics Statement

This study was approved by the Ethics Committee of the Eye & ENT Hospital of Fudan University (Approval No. 2019095‐2). Written informed consent was waived by the Institutional Review Board.

Conflicts of Interest

The authors of this manuscript declare no relationship with any companies whose products or services may be related to the subject matter of the article.

Gong W.‐T., Liu X.‐L., Chen X., et al., “Multi‐Sequence Fusion MRI Deep Learning Models for Discrimination of Recurrent Nasopharyngeal and Osteoradionecrosis,” World Journal of Otorhinolaryngology ‐ Head and Neck Surgery 0 (2026): 1‐10. 10.1002/wjo2.70165.

Data Availability Statement

The data sets analyzed during the current study are not publicly available due to patient privacy and institutional policies. However, interested readers may contact the corresponding author to request access to de‐identified data, which may be shared on reasonable request and pending approval by the institutional ethics committee.