TFE3‐DualNet: An Interpretable Foundation Model‐Based Deep Learning Ensemble for Diagnosing TFE3‐Rearranged Renal Cell Carcinoma From Whole‐Slide Images in a Two‐Center Cohort

ABSTRACT

Background

TFE3‐rearranged renal cell carcinoma (TFE3‐rRCC) is a rare, aggressive subtype that predominantly affects adolescents and young adults. Its marked morphologic heterogeneity can delay recognition and downstream confirmatory testing.

Methods

We assembled a two‐center retrospective cohort of patients < 30 years with renal cell carcinoma (n = 228; 59 TFE3‐rRCC), using fluorescence in situ hybridization (FISH) as the reference standard. Model development was performed in a development cohort (n = 129), followed by independent external validation (n = 99). We developed TFE3‐DualNet, an ensemble of weakly supervised CLAM models trained on routine hematoxylin and eosin (H&E) whole‐slide images (WSIs) using patch embeddings extracted from two pathology foundation models (UNI and CHIEF). We compared performance with three immunohistochemistry (IHC) scoring methods and a feature‐fusion CLAM baseline using concatenated H&E‐derived UNI and CHIEF features, and assessed interpretability by attention mapping.

Results

In the external validation cohort, TFE3‐DualNet achieved an area under the receiver operating characteristic curve (AUROC) of 0.932, with accuracy 0.879, sensitivity 0.893, and specificity 0.873. The model outperformed IHC scoring methods (AUROC 0.793–0.819; all p < 0.05) and exceeded the feature‐fusion baseline (AUROC 0.906). Attention hotspots localized to diagnostically relevant tumor regions and showed concordance with TFE3 IHC patterns.

Conclusions

TFE3‐DualNet showed encouraging performance as an interpretable H&E WSI‐based screening model for TFE3‐rRCC in young patients, supporting its potential use to prioritize confirmatory testing and pathologist review in routine diagnostic workflows.

Abbreviations

AUPRC

area under the precision‐recall curve

AUROC

area under the receiver operating characteristic curve

CI

confidence interval

FISH

fluorescence in situ hybridization

H&E

hematoxylin and eosin

ICI

immune checkpoint inhibitor

IHC

immunohistochemistry

IQR

interquartile range

RCC

renal cell carcinoma

TFE3‐rRCC

TFE3‐rearranged renal cell carcinoma

WSI

whole‐slide image

1. Introduction

TFE3‐rearranged renal cell carcinoma (RCC), defined by Xp11.2 translocations involving the TFE3 gene, is a rare subtype of renal cell carcinoma that predominantly affects children and young adults but is increasingly recognized in older patients as well [1, 2]. Compared with other RCC subtypes, TFE3‐rRCC more frequently presents with advanced‐stage disease, aggressive clinical behavior, and unfavorable outcomes, making timely and accurate diagnosis crucial for optimizing treatment and follow‐up [3, 4].

In current clinical practice, the diagnosis of TFE3‐rRCC relies on an integrated workflow that combines morphologic assessment on hematoxylin and eosin–stained slides, TFE3 immunohistochemistry [5, 6], and confirmatory fluorescence in situ hybridization or RNA sequencing [7, 8]. However, TFE3 IHC is susceptible to pre‐analytic and analytic variability, including differences in antibodies, staining platforms, and scoring criteria, which can lead to both false‐positive and false‐negative results [9]. FISH and RNA‐based assays are considered reference standards but are labor‐intensive, costly, and technically demanding [10, 11], and therefore are not routinely performed for all RCC cases, particularly when TFE3‐rRCC is not initially suspected [12, 13]. As a result, a proportion of patients with TFE3‐rRCC may remain underdiagnosed or experience diagnostic delay.

Recent advances in computational pathology and deep learning have created new opportunities for extracting molecular and diagnostic information directly from hematoxylin and eosin (H&E)‐stained whole‐slide images (WSIs) [13, 14, 15, 16, 17, 18, 19]. Traditional machine‐learning approaches using hand‐crafted features have shown that TFE3‐rRCC can be distinguished from other RCC subtypes to some extent on H&E images, but these models are limited by small sample sizes, single‐center cohorts, and restricted generalizability [13]. More recently, pathology foundation models pretrained on large, heterogeneous WSI collections have demonstrated superior performance and transferability across multiple cancer types and tasks [14, 15, 16, 17, 18, 19]. To date, however, no study has systematically evaluated foundation model‐based deep learning for the diagnosis of TFE3‐rRCC or directly compared such models with established TFE3 IHC scoring methods.

Here, we developed and validated TFE3‐DualNet, an ensemble multiple‐instance learning model that integrates two state‐of‐the‐art pathology foundation models, UNI [14] and CHIEF [17], to identify TFE3‐rRCC using H&E‐stained WSIs from patients younger than 30 years. In a two‐center retrospective cohort of 228 RCC patients, we compared the diagnostic performance of TFE3‐DualNet with three widely used TFE3 IHC scoring methods and with an early feature fusion‐based clustering‐constrained attention multiple‐instance learning (CLAM) model using H&E‐derived UNI and CHIEF features. We further performed ensemble‐based uncertainty analysis and attention‐heatmap visualization to assess case‐level confidence and spatial concordance with TFE3 IHC. The overall study workflow and model development pipeline are illustrated in Figure 1. Our aim was to determine whether a foundation model‐based, interpretable deep learning system could serve as a sensitive, cost‐efficient screening tool for TFE3‐rRCC in routine clinical practice.

FIGURE 1.

Study design. (a) Patient cohort used for the SYSUCC development cohort and SYSUFAH external validation cohort. (b) Workflow for initial data preparation. (c) TFE3‐DualNet construction and training. Pathological features were extracted using UNI and CHIEF networks, and then modeled using CLAM with five Monte Carlo resampling repeats. For each backbone, the maximum positive‐class probability across the five repeats was obtained, and the two backbone‐specific probabilities were averaged to generate the final TFE3‐DualNet score. (d) Whole pipeline of TFE3‐DualNet model training and downstream analysis.

2. Materials and Methods

2.1. Participants

In conventional pathological practice, FISH is used to detect the translocation for TFE3 and TFEB genes using a break‐apart probe and is recommended for diagnosing renal cell carcinomas in patients under 30 years old [20]. This retrospective study screened consecutive patients younger than 30 years with pathologically diagnosed RCC who underwent surgical resection at the First Affiliated Hospital of Sun Yat‐sen University or Sun Yat‐sen University Cancer Center between 2010 and 2021. No formal sample‐size calculation was performed because the cohort was defined retrospectively by all patients meeting the prespecified eligibility criteria during the study period. Cases were eligible for final analysis when formalin‐fixed paraffin‐embedded (FFPE) tissue, H&E WSIs, interpretable TFE3 fluorescence in situ hybridization (FISH) results, and required clinicoradiologic data were available. Patients were excluded before final analysis if FFPE tissue was unavailable, tumor tissue was insufficient for representative analysis, FISH yielded no interpretable signal because of section quality or DNA degradation, or CT/MRI data were missing. According to these criteria, 273 young RCC patients were assessed, 45 were excluded, and 228 were included in the final analysis, comprising 129 patients in the SYSUCC development cohort and 99 patients in the SYSUFAH external validation cohort (Figure S1). Clinical information, including age, gender, stage, tumor size, and TNM was collected. The retrospective analysis of anonymized patient data was approved by the institutional review boards of the First Affiliated Hospital of Sun Yat‐sen University and Sun Yat‐sen University Cancer Center (No. 2023442), with a waiver of informed consent.

2.2. Data Acquisition

2.2.1. Fluorescence In Situ Hybridization Assay and Evaluation

FISH was conducted to confirm the diagnosis of TFE3‐rearranged RCC cases using a dual‐color break‐apart probe (GSP TFE3; Anbiping Company), as previously described [10]. Tumor sections of 4 μm thickness were deparaffinized and heated for 25 min at 100°C in distilled water. The sections were then incubated in a pepsin solution (4 mg/mL in 0.1 N HCl and 0.9% NaCl) at 37°C for 20 min for digestion. After a 3‐min wash in 2× SSC buffer, the sections were dehydrated in graded ethanol solutions (70%, 85%, and 100%) for 3 min each at room temperature. Following dehydration and drying, the probe was added to the tumor region on each slide. The slides were co‐denatured with the target DNA at 80°C for 5 min, then hybridized at 37°C overnight in a humidified incubator. Post‐hybridization, several washing steps were performed, and the sections were air‐dried. The sections were then counterstained with DAPI (In situs) and the cover slipped. The slides were examined using a Zeiss LSM880 microscope (Zeiss) and assessed according to previously reported methods [11]. A negative result appears as a fusion signal or closely approximated green‐red signal pattern, whereas a TFE3 rearrangement was indicated by a split‐signal pattern, where the green and red signals were separated by a distance greater than two signal diameters. At least 100 nonoverlapping tumor nuclei were examined for each case at 1000× magnification. A positive result was defined when at least 10% of the tumor nuclei showed a split‐signal pattern. Representative images of the FISH assay are shown in Figure S2. FISH assessment was performed before model development and was not informed by model outputs. FISH results were used as reference labels for model training, tuning, model selection, and performance evaluation.

2.2.2. Immunohistochemistry and TFE3 Nuclear Immunoreactivity Evaluation

Tumor sections of 4 μm thickness were used for immunohistochemistry (IHC) to assess TFE3 nuclear immunoreactivity. Each TFE3 slide was evaluated using three scoring systems [21, 22, 23], referred to as IHC method 1, 2, and 3 in this study.

The TFE3 expression in tumor samples was evaluated using three immunohistochemistry (IHC) scoring methods, with cytoplasmic staining ignored in scoring method 1 [16]. The intensity of nuclear immunoreactivity in all tumor nuclei was graded as negative (0), mild (1+), moderate (2+), or strong (3+). Cases were considered positive when the labeling intensity was moderate (2+) to strong (3+), whether the staining was focal or diffuse, and apparent at low‐power magnification (×40). Cases with focal staining in < 10% of cells and mild intensity labeling were excluded. In Scoring Method 2 [17], the percentage of stained tumor nuclei was assessed. The categories were negative (0), focal (1+) with < 10% stained nuclei, moderate (2+) with 10%–50% stained nuclei, and diffuse (3+) with more than 50% stained nuclei. Cases were considered positive when labeling showed moderate (2+) or diffuse (3+) immunoreactivity. Scoring Method 3 [18] was based on both nuclear staining intensity and the percentage of stained tumor cell nuclei. The score, termed “Xp11” was calculated by multiplying the nuclear intensity (where 0 = no staining, 1 = mild staining, 2 = moderate staining, 3 = strong staining) by the percentage of stained tumor cell nuclei (ranging from 0% to 100%). The scoring categories were defined as negative (0) with a score < 25, weak positive (1+) with a score of 26–100, moderate positive (2+) with a score of 101–200, and strong positive (3+) with a score of 201–300. The representative images are shown in Figure S3.

2.2.3. Whole‐Slide Image Collection and Preprocessing

A representative H&E‐stained tumor slide (the one containing the highest tumor grade in the specimen) was collected for each patient in the development and external validation cohort. Next, a WSI in the SVS format was generated for each representative slide using a KF‐PRO‐020 scanner (KFBIO, Ningbo, China) at 40× equivalent magnification (0.25 μm/pixel). The WSIs were then annotated and reviewed by expert pathologists using the ASAP platform. Patches of 256 × 256 pixels were generated at 10× resolution (approximately 1 μm/pixel) using OpenSlide and the annotated masks by creating a grid of potential tiles starting from the top‐left corner to the lower right of the WSIs, including areas outside the tumor regions. Candidate tiles with an overall tumor percentage calculated based on the masks of more than 50% (compared to the whole area in each tile) were selected. The Vahadane et al. [24] color normalization method was used on patches to erase the H&E stain difference from different centers.

2.3. Development of the Ensemble Deep Learning Model

Patients from Sun Yat‐sen University Cancer Center formed the SYSUCC development cohort (n = 129), and patients from the First Affiliated Hospital of Sun Yat‐sen University formed the SYSUFAH external validation cohort (n = 99). Model development was performed within the SYSUCC cohort using five Monte Carlo resampling repeats based on the CLAM training framework. In each repeat, the SYSUCC cases were internally partitioned into training, tuning, and internal‐validation subsets. The training subset was used for model fitting, the tuning subset was used for loss monitoring, early stopping, and model selection, and the internal‐validation subset was used to assess repeat‐specific internal performance. Because these internal subsets varied across repeats, they were not treated as fixed reporting cohorts. Development cohort performance was summarized at the patient level for all 129 SYSUCC cases, and the SYSUFAH cohort served as the only fully independent external validation cohort. Among the five candidate foundation models: UNI [14], CHIEF [17], CONCH [15], Virchow [16] and Prov‐GigaPath [18], UNI and CHIEF were selected for their superior AUROC performance (Figure S4). The pretrained weights of each model were utilized for feature extraction from all WSIs in this study. Before feature extraction, image patches were preprocessed using Torchvision transforms: resized to 224 × 224 pixels, converted to tensors, and normalized using the mean and standard deviation values from the ImageNet dataset (mean = [0.485, 0.456, 0.406], std = [0.229, 0.224, 0.225]). This normalization, commonly referred to as ImageNet normalization, aligns the data distribution with that of the pretraining datasets, enhancing the effectiveness of transfer learning.

The weakly supervised model CLAM [19], based on clustering‐constrained‐attention multiple‐instance learning, was trained for TFE3 classification. Two separate CLAM models were developed: one using features extracted from the UNI model (UNI‐based CLAM) and the other using features from the CHIEF model (CHIEF‐based CLAM).

During model development, a grid search strategy combined with five Monte Carlo resampling repeats was applied within the SYSUCC development cohort to optimize hyperparameters and improve model robustness (Figure S5). For each repeat, a CLAM model was fitted on the repeat‐specific training subset, the tuning subset was used to monitor validation loss/AUC, support early stopping, and guide model selection, and the internal‐validation subset was used to assess repeat‐specific internal performance. Because the internal partitions varied across repeats, they were used for model development and internal resampling assessment rather than as fixed independent reporting cohorts.

In addition to individual UNI‐based‐CLAM and CHIEF‐based CLAM models, a feature fusion‐based model by combining features extracted from both the UNI and CHIEF models prior to CLAM training was constructed. This early feature fusion involved concatenating the feature embeddings extracted from the two foundation models, aiming to integrate complementary H&E‐derived features from the two foundation models. The fused features were then fed into the same CLAM model. The same feature normalization steps, cross‐validation, and parameter search strategy were applied.

2.4. Model Evaluation

An ensemble model was created to achieve optimal classification performance by combining the UNI‐based CLAM and CHIEF‐based CLAM outputs. For each patient, the maximum positive‐class probability across five Monte Carlo repeats was first obtained separately for the CHIEF‐based and UNI‐based CLAM models. The two backbone‐specific maximum probabilities were then averaged to generate the final TFE3‐DualNet probability score. TFE3‐DualNet was evaluated using AUROC and the area under the precision‐recall curve (AUPRC). Binary predictions were generated by applying a threshold of 0.5 to the final probability score. Performance metrics, including accuracy, sensitivity, specificity, precision, F1 score, and negative predictive value (NPV), were calculated at the cohort level for the SYSUCC development cohort and the SYSUFAH external validation cohort. The SYSUFAH cohort was held out from model development and served as the only fully independent performance estimate.

Comparisons were made between TFE3‐DualNet and the three IHC scoring systems using all the aforementioned metrics. This comprehensive evaluation demonstrated how TFE3‐DualNet's performance compared with conventional diagnostic methods, demonstrating its potential utility in clinical practice.

2.5. Model Interpretability and Visualization

To interpret the relative importance of different regions within each slide for the final slide‐level prediction, we followed the attention‐based visualization strategy described by Lu et al. [19]. Specifically, for each model we extracted the unnormalized attention scores (before the softmax operation) from the attention branch corresponding to the predicted class and saved them for all patches within the slide. These raw attention scores were then converted to percentile ranks and linearly rescaled to the range 0–1, with 1.0 representing the highest attention and 0 the lowest. The normalized scores were mapped to RGB values using a diverging colormap and visualized at their corresponding spatial coordinates on the WSI. High‐attention regions, indicating strong contributions to the model's prediction, were shown in red, whereas low‐attention regions were shown in blue and intermediate values in cyan to yellow. The resulting region‐of‐interest heatmaps were overlaid onto the original H&E WSI with a transparency of 0.4, allowing simultaneous inspection of both the model‐derived attention and the underlying morphology.

For TFE3‐DualNet, an ensemble attention heatmap was generated by first computing attention maps separately from the CHIEF‐based CLAM and UNI‐based CLAM models and then averaging their normalized attention scores at each patch location, yielding a single integrated attention map for visualization.

2.6. Software

ASAP (https://computationalpathologygroup.github.io/ASAP) was used for annotation, OpenSlide (v.1.3.0) was used for patch generation. Python (v.3.8.16) was used for all experiments and analyses in the study. For model training, PyTorch (v.1.11.0) and Torchvision (v.0.14.1) were used. BioRender (www.biorender.com) was used for figure preparation. Implementations of feature extraction and downstream classification in the study were obtained from the following links: UNI (https://huggingface.co/MahmoodLab/UNI), CHIEF (https://github.com/hms‐dbmi/CHIEF), CONCH (https://huggingface.co/MahmoodLab/CONCH), Virchow (https://huggingface.co/paige‐ai/Virchow), Prov‐GigaPath (https://huggingface.co/prov‐gigapath/prov‐gigapath), CLAM (https://github.com/mahmoodlab/CLAM).

2.7. Hardware

Tiling and color normalization of WSIs were performed on Intel Xeon central processing units (CPUs). Feature extraction using pretrained pathological foundation models and CLAM model training were accelerated by NVIDIA V100 graphics processing units (GPUs) on Tianhe‐2 supercomputer nodes.

2.8. Statistical Analysis

Continuous clinicopathological variables are presented as medians with interquartile ranges, and categorical variables are presented as counts and percentages. Model performance was summarized at the patient level for the SYSUCC development cohort (n = 129; 31 TFE3‐rRCC and 98 non‐TFE3‐rRCC) and the SYSUFAH external validation cohort (n = 99; 28 TFE3‐rRCC and 71 non‐TFE3‐rRCC), with the SYSUFAH cohort serving as the only fully independent performance estimate. Two‐sided 95% confidence intervals were computed with 1000 patient‐level bootstrap iterations. TFE3‐DualNet performance was evaluated using receiver operating characteristic curves and precision‐recall curves. AUROC and AUPRC were calculated to compare model performance with IHC methods. Paired DeLong tests [25] were used for AUROC comparisons, and permutation tests [26] with 1000 label permutations were used for AUPRC comparisons. Binary metrics, including accuracy, sensitivity, specificity, precision, F1 score, and NPV, were computed using a threshold of 0.5 applied to the final TFE3‐DualNet probability score. A two‐sided alpha level of 0.05 was used, and no multiplicity adjustment was applied. Analyses were performed in Python (v.3.8.16) using NumPy, pandas, scikit‐learn, and SciPy.

3. Results

3.1. Patient Characteristics

A total of 228 patients were included in this retrospective study, comprising 59 patients (25.9%) with FISH‐confirmed TFE3‐rRCC and 169 patients (74.1%) with non‐TFE3‐rRCC. Patients with TFE3‐rRCC were generally younger, with a median age of 22.0 years (interquartile range [IQR], 12.0–25.0), compared to 27.0 years (IQR, 24.0–29.0) for non‐TFE3‐rRCC patients, and showed a female predominance (59.3% vs. 31.4%). Clinically, TFE3‐rRCC patients presented with more aggressive disease characteristics, including higher tumor grades (WHO/ISUP grade 3 or 4 in 55.9% vs. 42.6% in non‐TFE3‐rRCC), advanced pathological stages (stage III or IV in 57.6% vs. 13.0%), larger median tumor size (5.5 cm [IQR, 3.6–8.0 cm] vs. 4.0 cm [IQR, 2.9–6.0 cm]), and higher rates of lymph node metastases (27.1% vs. 5.9%) and distant metastases (10.2% vs. 4.7%). These findings suggest that TFE3‐rRCC is a more clinically aggressive subtype compared to other RCCs. The final cohort comprised a SYSUCC development cohort of 129 patients from Sun Yat‐sen University Cancer Center (31 TFE3‐rRCC and 98 non‐TFE3‐rRCC) and a SYSUFAH external validation cohort of 99 patients from the First Affiliated Hospital of Sun Yat‐sen University (28 TFE3‐rRCC and 71 non‐TFE3‐rRCC). Both cohorts were staged using the TNM (2017) system and graded according to the WHO/ISUP criteria. Detailed clinical features of patients in the development and the external validation cohorts are described in Table 1. Participant screening and center‐specific exclusion reasons are summarized in Figure S1.

TABLE 1.

Clinicopathological characteristics of patients in the SYSUCC development cohort and SYSUFAH external validation cohort.

	All (n = 228)		Development cohort (n = 129)		External validation cohort (n = 99)
	Non‐TFE3 (n = 169)	TFE3 (n = 59)	Non‐TFE3 (n = 98)	TFE3 (n = 31)	Non‐TFE3 (n = 71)	TFE3 (n = 28)
Sex
Male	116 (68.6%)	24 (40.7%)	63 (64.3%)	11 (35.5%)	53 (74.6%)	13 (46.4%)
Female	53 (31.4%)	35 (59.3%)	35 (35.7%)	20 (64.5%)	18 (25.4%)	15 (53.6%)
Age, years	27.0 (24.0–29.0)	22.0 (12.0–25.0)	27.0 (25.0–29.0)	22 (18.5–29.0)	27 (24.0–29.0)	19.0 (11.8–25.3)
WHO/ISUP grade
1	21 (12.43%)	0 (0.00%)	10 (10.20%)	0 (0.00%)	11 (15.49%)	0 (0.0%)
2	76 (44.97%)	26 (44.07%)	44 (44.90%)	11 (35.48%)	32 (45.07%)	15 (53.6%)
3	64 (37.87%)	30 (50.85%)	36 (36.73%)	19 (61.29%)	28 (39.44%)	11 (39.29%)
4	8 (4.73%)	3 (5.08%)	8 (8.16%)	1 (3.23%)	0 (0.00%)	2 (7.14%)
Pathologic stage
I	120 (71.01%)	20 (33.9%)	66 (67.35%)	13 (41.94%)	54 (76.06%)	7 (25.0%)
II	27 (15.98%)	5 (8.5%)	18 (18.37%)	3 (9.68%)	9 (12.68%)	2 (7.1%)
III	12 (7.1%)	21 (35.59%)	6 (6.12%)	9 (29.03%)	6 (8.45%)	12 (42.86%)
IV	10 (5.92%)	13 (22.03%)	8 (8.16%)	6 (19.35%)	2 (2.82%)	7 (25.0%)
pT
T1	129 (76.33%)	33 (55.93%)	68 (69.39%)	18 (58.06%)	61 (85.92%)	15 (53.57%)
T2	28 (16.57%)	11 (18.64%)	20 (20.41%)	5 (16.13%)	8 (11.27%)	6 (21.43%)
T3	9 (5.33%)	12 (20.34%)	9 (9.18%)	8 (25.81%)	0 (0.00%)	4 (14.29%)
T4	3 (1.78%)	3 (5.08%)	1 (1.02%)	0 (0.00%)	2 (2.82%)	3 (10.71%)
pN
N0	159 (94.08%)	43 (72.88%)	90 (91.84%)	24 (77.42%)	69 (97.18%)	19 (67.86%)
N1	10 (5.92%)	16 (27.12%)	8 (8.16%)	7 (22.58%)	2 (2.82%)	9 (32.14%)
M
M0	161 (95.27%)	53 (89.83%)	91 (92.86%)	29 (93.55%)	70 (98.59%)	24 (85.71%)
M1	8 (4.73%)	6 (10.17%)	7 (7.14%)	2 (6.45%)	1 (1.41%)	4 (14.29%)
Tumor size, cm	4.0 (2.9–6.0)	5.5 (3.6–8.0)	4.6 (2.9–7.5)	5.5 (3.6–7.3)	4.0 (2.9–4.9)	5.5 (3.6–8.8)

Note: Data are n (%) or median (IQR). Pathologic stage was classified according to the TNM 2017 staging system.

Abbreviations: ISUP, International Society of Urological Pathology; pN, regional lymph node status; pT, pathological T stage; WHO, World Health Organization.

The TFE3‐DualNet was constructed by integrating the CHIEF‐based and UNI‐based CLAM models, and its performance was compared against three IHC scoring systems [21, 22, 23]. Notably, TFE3‐DualNet demonstrated excellent predictive accuracy for TFE3 status, with an area under the receiver operating characteristic curve (AUROC) of 0.996 (95% CI: 0.987–1.000) in the SYSUCC development cohort and 0.932 (95% CI: 0.868–0.978) in the SYSUFAH external validation cohort. The results of the DeLong test revealed statistically significant differences (p value ranging from 0.024 to 0.046) between TFE3‐DualNet and each of the three IHC scoring systems, supporting the diagnostic performance of TFE3‐DualNet in predicting TFE3 rearrangement status in renal cell carcinoma patients (Figure 2a).

FIGURE 2.

Performance of TFE3‐DualNet and IHC methods. (a) ROC curves for TFE3‐DualNet and IHC methods in the SYSUCC development cohort (n = 129) and SYSUFAH external validation cohort (n = 99). (b) Precision‐recall curves for TFE3‐DualNet and IHC methods in both cohorts. The positive class was TFE3‐rRCC, and FISH was used as the reference standard. Two‐sided 95% CIs were computed using 1000 patient‐level bootstrap iterations. p values for AUROC comparisons were calculated using paired DeLong tests, and p values for AUPRC comparisons were calculated using permutation tests with 1000 label permutations. No multiplicity adjustment was applied. IHC_M, IHC scoring method.

The feature fusion‐based CLAM model, trained on the concatenated features from UNI and CHIEF, achieved an AUROC of 0.978 (95% CI: 0.954–0.995) in the development cohort and 0.906 (95% CI: 0.834–0.962) in the external validation cohort. Compared to TFE3‐DualNet, which combined UNI‐based and CHIEF‐based CLAM predictions at a later stage, the feature fusion‐based model had a lower AUROC and did not significantly outperform conventional IHC scoring systems (p value ranging from 0.087 to 0.158). This finding suggests the feature fusion‐based model's comparatively modest predictive ability in identifying TFE3‐rRCC and highlights the superiority of late ensemble techniques over early feature fusion (Figure S6).

Given the imbalance in the dataset, in which TFE3‐positive patients represent approximately 25.9% of the total, the precision‐recall curves were used to evaluate TFE3‐DualNet's ability to classify the minority positive class without being influenced by the prevalence of negative cases. The high average precision scores, 0.987 (95% CI: 0.963–1.000) in the development cohort and 0.846 (95% CI: 0.684–0.960) in the external validation cohort, support the ability of TFE3‐DualNet to identify the minority TFE3‐positive patients. This indicates that TFE3‐DualNet achieves a high true‐positive rate while maintaining a low false‐positive rate. A permutation test further demonstrated a statistically substantial difference (p < 0.001) between TFE3‐DualNet and IHC scoring methods, exhibiting the model's superior predictive accuracy in identifying TFE3‐positive patients (Figure 2b).

Additional metrics, including accuracy, sensitivity, specificity, F1 score, precision, and NPV, were also calculated, with the results summarized in Table 2. In the SYSUCC development cohort, TFE3‐DualNet significantly outperformed the IHC scoring methods across all metrics. In the external validation cohort, TFE3‐DualNet achieved an accuracy of 0.879, surpassing IHC methods (ranging from 0.848 to 0.859). It also achieved markedly higher sensitivity (0.893 vs. IHC methods ranging from 0.643 to 0.750), underscoring its effectiveness in detecting TFE3‐positive patients. Although IHC method 1 exhibited slightly higher specificity (0.944) and precision (0.818), TFE3‐DualNet's superior F1 score (0.806 vs. 0.720–0.741) and NPV (0.954 vs. 0.870–0.900) highlight its overall better performance. These results support TFE3‐DualNet as a sensitive screening tool for identifying TFE3‐positive patients in this highly imbalanced cohort (Figure S7).

TABLE 2.

Diagnostic performance of TFE3‐DualNet and IHC scoring methods in the SYSUCC development cohort and SYSUFAH external validation cohort.

Dataset	Method	Accuracy	Sensitivity	Specificity	F1 score	Precision	NPV
Development cohort	Ensemble	0.922	1.000	0.898	0.861	0.756	1.000
	IHC_M1	0.806	0.839	0.796	0.675	0.565	0.940
	IHC_M2	0.698	0.968	0.612	0.606	0.441	0.984
	IHC_M3	0.767	0.935	0.714	0.659	0.509	0.972
External validation cohort	Ensemble	0.879	0.893	0.873	0.806	0.735	0.954
	IHC_M1	0.859	0.643	0.944	0.720	0.818	0.870
	IHC_M2	0.848	0.750	0.887	0.737	0.724	0.900
	IHC_M3	0.859	0.714	0.915	0.741	0.769	0.890

Note: Bold values indicate the highest value for each performance metric within each cohort.

Abbreviations: Ensemble, TFE3‐DualNet; IHC_M, IHC scoring method; NPV, negative predictive value.

3.2. Uncertainty Analysis of TFE3‐DualNet Predictions

To further investigate the confidence of TFE3‐DualNet, we examined the dispersion of slide‐level probabilities across the 10 ensemble members. Thirty representative cases from the SYSUCC development cohort and 30 cases from the SYSUFAH external validation cohort are shown in Figure 3. Corresponding uncertainty plots for the single‐backbone CLAM models are provided in Figure S8. In both cohorts, most correctly classified TFE3‐positive and TFE3‐negative slides exhibited mean probabilities far from the decision threshold and relatively small error bars, indicating high model confidence. In contrast, misclassified or borderline cases tended to cluster around the decision threshold and showed visibly larger variability across ensemble members, reflecting higher prediction uncertainty. These patterns suggest that TFE3‐DualNet can provide useful case‐level uncertainty estimates and may help flag challenging slides that warrant confirmatory IHC or FISH evaluation in clinical practice.

FIGURE 3.

Uncertainty analysis of TFE3‐DualNet predictions. Representative 30 cases from (a) the development cohort and (b) the external validation cohort are shown. Each dot represents the mean positive‐class probability across the 10 ensemble members (five UNI‐based CLAM models and five CHIEF‐based CLAM models) for a single slide, and the horizontal bars indicate ±1 standard deviation. Blue and red markers denote non‐TFE3‐rRCC and TFE3‐rRCC, respectively. Cases with predictions far from the decision threshold (0.5) show low variability, whereas borderline or misclassified cases tend to cluster around the threshold with higher uncertainty.

3.3. Model Interpretability Analysis

The weakly supervised CLAM framework was used to assess whether model predictions were driven by histologically meaningful regions that align with TFE3‐positive IHC staining. The CLAM model makes slide‐level predictions by first identifying and aggregating regions in the WSI that hold high diagnostic importance (high‐attention score) while ignoring regions of low diagnostic relevance (low‐attention score) [19]. Based on this approach, attention heatmaps for TFE3‐DualNet were generated to visualize and interpret the relative importance of each region in the WSI for TFE3‐rRCC patients. Notably, regions most strongly attended to by the models (highlighted in red on the attention heatmap) on H&E‐stained WSI showed spatial concordance with TFE3‐positive regions on IHC staining images (Figure 4a,b). The top eight highly attended patches from each foundation model‐based CLAM, all exhibiting typical tumor features from different regions, further illustrate the meaningful differences between the foundation‐based CLAM models and their consistent concordance with IHC results (Figure 4c). Similar concordant patterns between TFE3‐DualNet attention and TFE3 IHC were observed in multiple additional TFE3‐positive cases (Figure S9). In each case, CHIEF‐based CLAM and UNI‐based CLAM produced partially overlapping but not identical attention distributions, with certain TFE3‐positive tumor foci highlighted more strongly by one model than the other. The TFE3‐DualNet ensemble, which averages the normalized attention scores from both models, generated more stable and spatially coherent high‐attention regions that closely followed the extent of TFE3 staining on the corresponding IHC slides. These cross‐case visualizations suggest that the ensemble model effectively integrates complementary information from the two foundation model‐based CLAM networks and is less likely to overlook true TFE3‐positive tumor regions, which may reduce the risk of false‐negative predictions. Furthermore, these attention‐based visualizations may help illustrate morphological regions associated with model predictions.

FIGURE 4.

Spatial concordance between TFE3‐DualNet attention and TFE3 immunohistochemistry in a TFE3‐rRCC case. (a) Whole‐slide overview of a representative TFE3‐rRCC case. From left to right: H&E‐stained WSI, TFE3 IHC WSI, and TFE3‐DualNet attention heatmap overlaid on the H&E image. Purple and orange boxes indicate two representative tumor regions. The attention heatmap is displayed using a blue‐to‐red colormap, with red corresponding to the highest attention scores. (b) Magnified views of the boxed regions showing H&E (left), TFE3 IHC (middle), and the corresponding TFE3‐DualNet attention map (right). In both regions, areas with the highest model attention (red) co‐localize with TFE3‐positive tumor foci on IHC, whereas nondiagnostic background tissue receives low attention. (c) Top 16 high‐attention H&E patches, consisting of the top eight patches from the CHIEF‐based CLAM and the top eight patches from the UNI‐based CLAM, selected from the same case. Scale bars: 2.5 mm in whole‐slide overview panels, 625 μm in magnified‐region panels, and 200 μm in high‐attention patch panels.

3.4. Challenging Cases Analysis

Challenging cases in the external validation cohort were closely examined. Among the nine patients predicted as positive by TFE3‐DualNet in the SYSUFAH external validation cohort (Figure 5a), two were diagnosed as positive by all IHC scoring methods (Figure 6a and Figure S10b), raising the possibility that these discordant cases may reflect limitations of the reference assay or sampling in individual cases. Further RNA sequencing would be required for definitive molecular adjudication.

FIGURE 5.

Confusion matrices of TFE3‐DualNet and IHC scoring methods in the external validation cohort. (a) Confusion matrix of TFE3‐DualNet in the external validation cohort, showing three false‐negative predictions for TFE3‐rRCC. (b) Confusion matrix of IHC scoring method one in the external validation cohort. (c) Confusion matrix of IHC scoring method two in external validation cohort. (d) Confusion matrix of IHC scoring method three in the external validation cohort. Positive: TFE3‐rRCC diagnosis; Negative: non‐TFE3‐rRCC diagnosis. IHC_M, IHC scoring method.

FIGURE 6.

Representative challenging cases with discrepant TFE3‐DualNet and reference‐standard labels in the external validation cohort. For each case, the columns show the H&E WSI, TFE3 IHC WSI, and TFE3‐DualNet attention heatmap; boxed regions correspond to higher‐magnification H&E and IHC views on the right. (a) Case labeled as non‐TFE3‐rRCC by FISH but predicted as TFE3‐rRCC by TFE3‐DualNet and all IHC scoring methods, with strong model attention concentrated in a tumor region showing nuclear TFE3 staining. (b) TFE3‐rRCC case (FISH‐ and IHC‐positive) misclassified as non‐TFE3‐rRCC by TFE3‐DualNet, in which TFE3‐positive tumor foci receive relatively low attention, while necrotic, fibrotic, or other nondiagnostic tissue is preferentially highlighted. Scale bars: 2.5 mm in whole‐slide overview panels, and 625 μm in magnified‐region panels.

By contrast, the three TFE3‐rRCC patients misdiagnosed as negative by TFE3‐DualNet (Figure 5a), while correctly diagnosed by the IHC scoring methods, were in early pathological stages and exhibited typical clear cell RCC or papillary RCC patterns. In these cases, high‐attention regions of TFE3‐DualNet were mainly located in necrotic, fibrotic, or other nondiagnostic areas, whereas the TFE3‐positive tumor foci attracted only limited attention (Figure 6b and Figure S10). Cases like these are often misdiagnosed in conventional diagnosis when relying solely on H&E images, highlighting the need for further studies on diagnostic models that utilize multisource or multimodal data.

Despite the false predictions in the external validation cohort, TFE3‐DualNet correctly diagnosed all the TFE3‐rRCC patients who were falsely diagnosed by all three IHC scoring methods (Figure 5b–d), indicating that the model provides complementary diagnostic value to IHC and may serve as a robust screening tool for initial evaluation in routine clinical practice.

4. Discussion

In this two‐center retrospective study, we developed and externally validated TFE3‐DualNet, a pathology foundation model‐based ensemble for identifying TFE3‐rRCC on H&E‐stained WSIs from young RCC patients. Across the SYSUCC development and SYSUFAH external validation cohorts, TFE3‐DualNet consistently outperformed three established TFE3 IHC scoring methods and a feature fusion‐based CLAM model, achieving AUROCs of 0.996 and 0.932, respectively, and maintaining high sensitivity while preserving reasonable specificity. By operating directly on routinely available H&E slides, TFE3‐DualNet addresses several practical limitations of current IHC‐ and FISH‐based diagnostic workflows [9, 10, 11, 12].

Our findings extend previous work [13] using traditional machine‐learning approaches to distinguish TFE3‐rRCC from clear cell RCC on H&E images by leveraging pathology foundation models [14, 17] trained on large, diverse WSI datasets. Unlike hand‐crafted feature pipelines, foundation models such as UNI [14] and CHIEF [17] learn rich hierarchical representations of tumor morphology that transfer well across cancer types and institutions. Because UNI‐ and CHIEF‐based CLAM models emphasize partially different image characteristics, they attend to complementary regions within the same slide. The superior performance of TFE3‐DualNet over the feature fusion‐based model suggests that late‐stage ensembling of model predictions can better exploit this complementarity than simply concatenating feature embeddings at the input stage.

Beyond overall accuracy, TFE3‐DualNet provides case‐level uncertainty estimates and intuitive visual explanations that are directly relevant to clinical decision‐making. Correctly classified cases typically showed mean ensemble probabilities far from the decision threshold with narrow error bars, whereas misclassified or borderline cases clustered around the threshold and exhibited higher dispersion, indicating that such slides may warrant confirmatory IHC or FISH evaluation. Attention heatmaps displayed strong spatial concordance with TFE3 IHC staining across representative cases, and challenging‐case analysis revealed two patients labeled as negative by FISH but positive by all three IHC scoring methods and by TFE3‐DualNet, raising the possibility of false‐negative FISH results [10, 21]. Conversely, three early‐stage TFE3‐rRCC cases with deceptively clear cell or papillary morphology were misclassified as negative by TFE3‐DualNet, with attention maps focusing on necrotic, fibrotic, or otherwise nondiagnostic regions [12, 23]. Taken together, these observations suggest that TFE3‐DualNet is best positioned as a sensitive, interpretable screening tool that complements, rather than replaces, existing IHC/FISH‐based diagnostics.

This study has several strengths, including the use of development and external validation cohorts, systematic comparison against three TFE3 IHC scoring systems, incorporation of state‐of‐the‐art pathology foundation models [14, 17], and extensive interpretability and uncertainty analyses. Nonetheless, important limitations remain. First, despite including cases from two centers, the absolute number of TFE3‐rRCC patients is still limited, reflecting the rarity of this entity [1] and restricting the precision of performance estimates. Second, our cohort was restricted to patients younger than 30 years [3], and additional validation is required in older populations and in more diverse clinical settings.

Third, TFE3‐DualNet was developed as an H&E WSI‐based screening model. This design reflects the clinical challenge that TFE3‐rRCC can show variable morphology and overlap with other RCC subtypes, making definitive diagnosis on H&E morphology alone unreliable [12]. In routine practice, diagnosis relies on ancillary testing such as TFE3 IHC, FISH, or molecular assays [9, 10, 11, 12]. Expert survey and recent diagnostic‐series evidence further support this point, showing that many genitourinary pathologists require FISH confirmation and that TFE3/TFEB‐altered renal tumors can substantially overlap with common renal tumors [27, 28, 29]. Accordingly, a meaningful pathologist comparison would need to be framed as a triage task, in which readers record their suspicion for TFE3‐rRCC or their recommendation for ancillary TFE3 testing, rather than as a definitive H&E‐only diagnosis. Such a prespecified reader study was beyond the scope of this retrospective analysis and should be evaluated in future work. Future work should therefore explore multimodal extensions of TFE3‐DualNet that integrate clinical variables, molecular data, and either real or virtually generated IHC staining, as recent studies have shown that combining H&E with IHC or other modalities can substantially improve diagnostic performance [30, 31]. Prospective evaluation in larger, geographically diverse cohorts will also be essential to refine decision thresholds, uncertainty‐based triage strategies, and workflow integration. Despite these limitations, our results suggest that a foundation model‐based ensemble applied to routine H&E slides may support the detection of TFE3‐rRCC and help prioritize young RCC patients with suspected TFE3 rearrangements for confirmatory testing and timely access to appropriate systemic therapies and clinical trials [4, 32].

Author Contributions

Yu‐Hang Chen: methodology, software, data curation, investigation, validation, formal analysis, funding acquisition, writing – review and editing, writing – original draft, visualization. Quan‐Hui Xu: investigation, writing – review and editing, methodology, software, formal analysis, writing – original draft, data curation. Hao‐Hua Yao: software, formal analysis, investigation, data curation, methodology, writing – review and editing, writing – original draft. Ke‐Zhi Liu: writing – original draft, writing – review and editing, data curation, methodology, software, formal analysis. Cheng‐Peng Gui: writing – review and editing, data curation, investigation. Liang‐Min Fu: data curation, investigation, writing – review and editing. Ying‐Han Wang: data curation, investigation, writing – review and editing. Jiang‐Quan Zhu: data curation, investigation, writing – review and editing. Jun‐Cai Li: investigation, data curation, writing – review and editing. Min‐Yu Chen: data curation, investigation, writing – review and editing. Kang‐Bo Huang: data curation, investigation, writing – review and editing. Han‐Sen Lin: data curation, funding acquisition, investigation, writing – review and editing. Bing Liao: data curation, investigation, validation, writing – review and editing. Yun Cao: investigation, data curation, validation, writing – review and editing. Jin‐Huan Wei: data curation, investigation, supervision, writing – review and editing. Peng‐Ju Li: data curation, writing – review and editing, writing – original draft, supervision. Jun‐Hang Luo: writing – original draft, project administration, data curation, resources, funding acquisition, supervision, writing – review and editing. Jia‐Zheng Cao: conceptualization, data curation, supervision, funding acquisition, writing – review and editing, writing – original draft, project administration.

Funding

This study was supported by the Guangzhou Science and Technology Program (Key Research and Development Project, No. 2025B03J0091), the National Natural Science Foundation of China (No. 82373433 and No. 82573729), the Tianjin Health Science and Technology Project (No. TJWJ2026QN015), the Shenzhen Medical Research Fund (No. A2502007), the Guangdong Basic and Applied Basic Research Foundation (Guangdong‐Jiangmen Joint Fund, No. 2025A1515140247), and the Tianjin Key Medical Discipline Construction Project (No. TJYXZDXK‐3‐003A).

Ethics Statement

This study was conducted in accordance with the ethical standards of the Declaration of Helsinki, and the retrospective analysis of anonymous patient data was approved by the institutional review board of the First Affiliated Hospital of Sun Yat‐sen University and the Cancer Center Affiliated to Sun Yat‐sen University (no. 2023442), with a waiver of informed consent for retrospective analysis of anonymized data.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Figure S1: STARD participant flow diagram. The diagram shows eligibility assessment, center‐specific exclusions, availability of H&E whole‐slide images, interpretable TFE3 FISH reference‐standard results, and final analyzed numbers for the SYSUCC development cohort and SYSUFAH external validation cohort. FFPE, formalin‐fixed paraffin‐embedded; FISH, fluorescence in situ hybridization; RCC, renal cell carcinoma; WSI, whole‐slide image.

Figure S2: Representative images of the TFE3 break‐apart probe assay. (a) Negative result in a male patient showing fused or closely approximated green‐red signals. (b) Positive result in a male patient showing a split‐signal pattern. (c) Negative result in a female patient showing paired fused signals. (d) Positive result in a female patient showing paired split signals. FISH sections were examined using a Zeiss LSM880 microscope (Zeiss) at 1000× magnification.

Figure S3: Representative immunohistochemistry staining for TFE3 based on three scoring systems. (a–d) Nuclear staining intensity grades 0, 1+, 2+, and 3+ are shown according to scoring method 1; only 2+ and 3+ were considered positive. (e–h) Percentages of stained tumor nuclei scored as 0, 1+, 2+, and 3+ are shown according to scoring method 2; only 2+ and 3+ were considered positive. (i–l) Immunostaining scores of 0, 1+, 2+, and 3+ are shown according to scoring method 3; 1+, 2+, and 3+ were considered positive. Scale bar: 50 μm in all panels.

Figure S4: Selection of the optimal foundation models. Illustration of the ROC curves and AUROCs are shown for each foundation model and compared with the IHC scoring methods. UNI and CHIEF showed higher AUCs than CONCH, Virchow, and Prov‐GigaPath, with both AUROCs above 0.91. GIGA: Prov‐GigaPath.

Figure S5: Loss and accuracy curves of foundation‐based CLAM models. (a, b) Training accuracy and loss across five Monte Carlo repeats for the CHIEF‐based and UNI‐based CLAM models, respectively.

Figure S6: Diagnostic performance of the feature‐fusion model. (a) ROC curves of the feature‐fusion model and the three individual IHC models (M1–M3) in the development cohort (left) and the external validation cohort (right). (b) Confusion matrices of the feature‐fusion model in the development cohort (left) and the external validation cohort (right).

Figure S7: Performance comparison in different metrics across models. Accuracy, precision, recall, specificity, F1 score, and NPV are shown for the CHIEF‐based CLAM, UNI‐based CLAM, and TFE3‐DualNet in the external validation cohort. Red bars represent TFE3‐DualNet, blue bars represent CHIEF‐based CLAM, and green bars represent UNI‐based CLAM.

Figure S8: Uncertainty analysis of the UNI‐based and CHIEF‐based CLAM models. Representative 30 cases from the development cohort (a) and external validation cohort (b) are shown for CHIEF‐based and UNI‐based CLAM. Compared with the ensemble model TFE3‐DualNet (Figure 3), the single‐backbone CLAM models show greater prediction variability for several borderline cases, highlighting the uncertainty‐reducing effect of late‐stage ensembling.

Figure S9: Attention maps of CHIEF‐based CLAM, UNI‐based CLAM, and TFE3‐DualNet compared with TFE3 IHC in four TFE3‐rRCC cases. For each case, columns show the H&E WSI, attention maps of the three models and the corresponding TFE3 IHC WSI. TFE3‐DualNet generates more spatially coherent high‐attention regions that better align with TFE3‐positive tumor areas than the single‐backbone models. Scale bar: 2.5 mm in whole‐slide overview panels.

Figure S10: Representative misclassified cases by TFE3‐DualNet. (a) Two TFE3‐positive cases incorrectly predicted as negative (false negatives). (b) One TFE3‐negative case incorrectly predicted as positive (false positive). For each case, the panels show, from left to right, the H&E‐stained WSI, the corresponding TFE3 IHC WSI, and the TFE3‐DualNet attention heatmap. Scale bars: 2.5 mm in whole‐slide overview slides.

Table S1: STARD‐2015 checklist.

Data Availability Statement

WSI data will be made available to interested research partners upon reasonable request to the corresponding author, J.‐H.L., or the Institute of Precision Medicine at the First Affiliated Hospital, Sun Yat‐sen University in Guangzhou, China. Approval of a data transfer agreement by the legal departments of both the requesting and contributing institutions and ethical clearance are required for data sharing. Code used in this study is available at https://github.com/ChenYuhang243/TFE3‐rRCC‐Diagnosis and archived at Zenodo: https://doi.org/10.5281/zenodo.13893259.