This study included patients who were preoperatively diagnosed with glioma and underwent surgical treatment at the Department of Neurosurgery, the First Affiliated Hospital of Xi’an Jiaotong University, between October 2019 and October 2024. A total of 506 patients with histopathologically confirmed glioma and available genetic testing were initially screened according to the 2021 WHO classification of tumors of the central nervous system, 5th edition. After applying the inclusion and exclusion criteria, 345 patients who met all requirements were ultimately enrolled in the study cohort, as shown in Figure 1. The IDH status of all data was determined by DNA sequencing.

Inclusion Criteria:

Exclusion Criteria:

All patients underwent preoperative cranial MRI examinations and provided written informed consent prior to imaging. The study was approved by the Ethics Committee of the First Affiliated Hospital of Xi’an Jiaotong University and was registered at ClinicalTrials.gov (registration number: NCT05019196).

Given a complete histopathological image slice X, our goal is to predict the image label Y by analyzing the features extracted from discrimination patches x1,x2,…,xn. To achieve this, we designed a two-stage framework for the classification of IDH status, as illustrated in Figure 2. In the first stage, we encode the spatial features of the tumor. In the second stage, we introduce a novel HAB-MIL network, which comprises multiple layers of convolutional neural networks and an attention mechanism module. This module aggregates the features of the candidate instance in the Yimage-level prediction by recalibrating the importance coefficients for each instance.

In MIL, the data is organized into “bags”, each containing multiple “instances”. The label assigned to a bag depends on the instances it contains. For a binary classification task, the bag is labeled positive if at least one instance within it is positive; otherwise, it is labeled negative. In the weakly supervised histopathology image classification problem, the dataset D={X1,X2,…,XN} consists of NMRI images, where each slide xihas label yi∈{0,1}, and the objective is to train a model to predict the labels of MRIs. This relationship is formally expressed as (Equation 1):

where Y is the bag label, yi is the label of instance, and is the number of instances in each bag. In our work, this assumption indicates that MRI is from a glioma patient if it involves at least one lesion. Based on the assumption, the empirical loss is formulated by (Equation 2):

where hf represents a labeling function induced an MIL scoring function f, and l(·)can be any loss function. The MIL process can be mathematically represented as follows (Equation 3):

where f is the feature extraction function, his the instance-level aggregation operator, g is predicts the bag-level label. The model processes three MRI modalities: T1w, T2w and Flair, which represented as a 3D tensor x∈RB×C×H×W×D, where Bis the batch size, Cis the number of channels and H×W×Dis the spatial resolution. The final layer of 3D fully CNN outputs a series of 3D feature maps Ffinal with the shape of H*×W*×S*×D, where H∗, W∗, S* and Drepresent the high, width, spatial, and feature dimension of 3D feature maps, respectively. The feature extraction process consists of multiple 3D convolutional layers, batch normalization, ReLU activation, max pooling, and dropout layers. Finally, feature map is formulated as (Equation 4):

where W0 is a 3D convolution kernels, with size 3×3×3, and the output channels are 32. ∗indicates the 3D convolution operation. Batch Normalization (BN) is used to stabilize training. σis the ReLU activation function.

The spatial location of a tumor often provides critical diagnostic cues (24, 25), since we proposed Collateral location encoding (CLE), as illustrated in Figure 2. P(x,y,z)represents the voxel coordinates of the image. Since p=(x,y,z)typically takes continuous values, employing a smooth nonlinear transformation like GELU allows for a more natural modeling of the relationships between coordinates, unlike ReLU, which introduces abrupt discontinuities. The location feature map F(p)or each position of pixel p=(x,y,z) is calculated as (Equation 5):

where a1,a2,b1,b2 are the learnable weights and biases of our neural positional encoding layers. The learned features F and location encoding F^from the previous feature extraction are first normalized to benefit the training and backpropagation process. Then features concatenated to be the input of this phase, which can be denoted as (Equation 6):

where ‖means the concatenation operation, Norm(F), Norm(F^) refer to the normalization of F, F^, H′={Hnorm‖H^norm}={[H1norm‖H^1norm],…,[Hknorm‖H^knorm]} Hknorm,H^knorm∈Rn. Then, using the self-attention mechanism, the importance of the features can be calculated by (Equations 7, 8):

where W,b denote the learnable weight and bias parameters. H′and O represent the input and output of the adaptive feature importance weighting mechanism based on the attention module. Softmax operation is applied to normalize the attention scores and resulting in an attention weight matrix A that maps each weight to the interval [0,1]and guarantees the sum of these mapped attention weights to be 1. ⊗denotes element-wise multiplication. Through this process, the input features can be weighted and assigned different importance, enabling the model to emphasize critical features and thereby enhance overall performance.

Attention mechanisms are well known for dynamically highlighting important parts of input data by assigning different weights, allowing the model to focus on the most critical information for the task at hand (26–28). To enhance the accuracy of learning attention weights, we propose Dynamic Gated Attention (DGA), which achieves superior performance. As illustrated in Figure 2, the DGA module integrates two different processing methods.

The first component is an attention representation that uses Tanh as the activation function. With an output range of [−1,1], Tanh provides both positive and negative activation signals, allowing the model to capture complex relationship features. In addition, a dropout function is incorporated to mitigate overfitting. The attention weights are defined as (Equation 9):

The second component utilizes a different attention representation. As illustrated in Figure 3, we propose using Learnable Singular Value Decomposition (LSVD) to reduce the dimensionality of the feature map (29), followed by a Sigmoid activation function, which creates a “switch-like” effect, indicating whether the model should attend to or ignore a particular feature. The algorithm flowchart is shown in Algorithm 1.

Fact1 The feature map Ai can be decomposed as Ai=αUΣ′αV, where αU∈Rn×ℜ(Ai) is an orthogonal matrix with the left singular vector of Ai as the column vector, where αVT∈Rn×ℜ(A) is an orthogonal matrix formed by the right singular vectors of Ai. The matrix ∑′∈Rℜ(Ai)×ℜ(Ai) is a diagonal matrix consisting of the singular values of Ai.

Proof of the Fact 1: According to the Singular Value Decomposition (SVD), any matrix M∈Rm×n admits a factorization of the form M=U∑VT, where σ1≥σ2≥…≥σr>0 are the singular values and r=ℜ(M) is the rank of M. Hence, for Ai∈Rm×n, the decomposition is as shown in (Equation 10):

Here, μi,νi∈Rn×1 are the singular vectors, r=ℜ(Ai)is the number of nonzero singular values. For clarity, we denote the left singular matrix (μ1…μr) as αU, the diagonal matrix diag(σ1,σ2,…,σr) as Σ′, and the right singular matrix (ν1…νr)T as αVT, yielding: Ai=αUΣ′αVT.

The remaining challenge is to determine αU,Σ',αVT. A natural approach is to leverage neural networks to learn these latent representations. Specifically, assuming ∑′is a trainable diagonal matrix defined as (Equation 11):

where wi∈R are trainable parameters in the neural network, which are initialized prior to training and subsequently optimized via backpropagation.

Then, we utilize two learnable layers of X to fit αU,αVT as shown in Equation 12, 13):

Among these, fU(·),fV(·) are projection layers to transform the n×d matrix Xinto matrices of shape n×r. In our implementation, we share the same set of two linear layers for fU(·),fV(·) by employing a common set of linear transformation parameters Wa∈Rd×r,ba∈R1×r. The transformation is defined as (Equation 14):

Formally, we denote Y={y1,y2,…,yn} by a bag of n deep instance, the attention- based MIL pooling is defined by as shown in (Equations 15, 16):

where aLSVD and aTanh are attention weights.

In this study, two distinct datasets were used for model development and evaluation. Both datasets extract the same features and do not specify whether the IDH-mutant are of the IDH1 or IDH2 type.

In-house data set: The first data set was acquired from the First Affiliated Hospital of Xi’an Jiaotong University and comprised 789 MRI samples of 344 subjects (incorporating FLAIR, T1w, and T2w sequences, all occurrences of “T1w” refer to T1 without contrast injection, and “T1c” refers to T1 with contrast injection.), as summarized in Table 1. All patients underwent surgical resection followed by genetic sequencing to confirm IDH status. For patients diagnosed before 2021, we performed a centralized histopathological review. Two board-certified neuropathologists, who specialize in brain tumor pathology, re-examined the archival hematoxylin–eosin and immunohistochemistry slides (including IDH1 R132H, ATRX, p53 and other markers when available) and reviewed the original molecular pathology reports. Based on these materials, tumors were reclassified according to the 2021 World Health Organization Classification of Tumors of the Central Nervous System (5th edition), using all available molecular markers. Importantly, no additional molecular testing (such as TERT, EGFR or chromosomal analyses) was performed because of the retrospective design and limited archival tissue in some case (30, 31).

Public data set: The second data set was obtained from The Cancer Imaging Archive (TCIA), which included 1485 MRI samples from 495 subjects (395 wild-type cases and 100 mutant-type cases). Each subject in the TCIA dataset provided Flair, T1w, and T2w modalities, together with documented IDH status. Both datasets underwent identical pre-processing and post-processing pipelines. The diagnosis of glioma for all patients was performed according to the 2021 WHO classification of tumors of the central nervous system, 5th edition (32). Specifically, MRI volumes were preserved in the DICOM format and standardized to a uniform dimension of 240×240×155. Standard intensity normalization and artifact removal procedures were applied, followed by data augmentation strategies such as color jitter and random affine transformations to mitigate overfitting. This unified preprocessing approach allowed for direct comparisons between the two datasets and a robust evaluation of the proposed HAB-MIL framework.

The framework was implemented using Pytorch2.5.1, and all models were trained with NVIDIA 3090 GPU with CUDA 11.8. We use 3D convolutional neural networks as the deep instance generator. We set the output shape H*×W*×S*×D of ψto be 2×2×2×64 according to cross-validation. The input shapes of the MRI slices are 240×240×155 We set training epoch T to 100 and batch size to 2. Data enhancement strategies included color jitter and random affine transformations. The network was trained using the ADAM optimizer with an initial learning rate 1e-4 and weight decay of 1e-5. In all experiments, 60% of the subjects was used for training, 20% for model selection and hyperparameter tuning, and the remaining 20% for testing. The dataset was augmented with random flipping, random affine transformation, intensity scaling, color jitter as was done to train the segmentation network. No scans from the same subject were included in both the training and the testing samples to ensure data independence and avoid potential information leakage. Five-fold cross-validation is performed within the training set, where the validation portion in each fold is applied for model selection and hyperparameter verification. Each experiment was repeated five times to ensure fair comparisons. Evaluation metrics included accuracy (Acc), area under the curve (AUC), sensitivity (Sens), specificity (Spec) and ROC curves.

A series of ablation studies on the TCIA data set was carried out to evaluate the effectiveness of the two modules (CLE and DGA) in HAB-MIL. Initially, use only the 3D convolution and MIL modules (denoted as Ablation-1) to evaluate the performance of the backbone. In the backbone, features can still be extracted from three concatenated conventional magnetic resonance sequences for the prediction of IDH. Next, add only the CLE module to the backbone (denoted Ablation-2), using positional encoding as a feature, and train the model. Next, add only the DGA module to the backbone (denoted Ablation-3). The results of the ablation experiments are presented in Table 2. Compared to backbone results, the incorporation of tumor location encoding improved the accuracy of IDH prediction by 21.8%. This suggests that positional encoding enables our HAB-MIL to learn location-specific features, which aids in predicting consistent differences across all image regions. Then, by adding the DGA module and selecting key instances, the model’s accuracy was further improved by 24%. Finally, combining all three modules results in the best prediction performance. Therefore, the CLE module can be used to explore the spatial relationships of pixels, and the DGA module helps select key instances, improving IDH prediction.

Table 3 lists classification results of various methods for the prediction of IDH. Among the various methods compared, the SPE can be used as a baseline to measure the validity of segmentation. We first ablate the effects of our CLE in the HAB-MIL for the test. As can be seen, Sinusoidal Positional Encoding (SPE) which simply concatenates the sinusoidal positional encoding of the pixel locations (x, y) values directly into the encoder stage leads to a slight performance improvement of 2% in most metrics compared to those without it (denoted as “w/o SPE”). Bello propose neural positional encoding (NPE) for depth estimation (33). However, our HAB-MIL shows substantial performance improvements in all metrics by incorporating our CLE. Specifically, the AUC increased by 28.8% and the ACC by 27.1% compared to the model w/o SPE. This improvement can be attributed to the learnable auxiliary positional encoding, which allows the model to more flexibly capture and leverage position-specific feature information. From the comparison between these methods, it can be observed that CLE is more helpful for IDH prediction, as it incorporates information that accurately describe location and morphological features of tumor and peritumoral edema.

To evaluate the effectiveness and feasibility of the DGA module within HAB-MIL, we considered three attention weighting methods: max pooling, average pooling, and AB-MIL approaches. These methods were compared against our proposed HAB-MIL model. Specifically, the max-pooling-based and average-pooling-based methods follow the traditional MIL assumption, where the final subject-level prediction is obtained from the most significant instance or the average of all instances within a bag, respectively. In contrast, the attention-pooling-based method leverages an attention mechanism to assign weights to each instance embedding, facilitating learning at the bag level. In contrast, HAB-MIL considers a dynamic attention mechanism to selectively focus on the most relevant instances, effectively minimizing the inclusion of irrelevant information. As shown in Table 4, our proposed HAB-MIL achieved the best overall performance, attaining an accuracy of 90.4%. Max-pooling demonstrated the lowest performance among all methods, with an AUC of 69.3%, which is expected as it only considers the single most prominent instance for prediction. In contrast, the mean-based approach, while accounting for all instances, introduces a significant amount of irrelevant information, leading to less satisfactory results. Moreover, although plain attention pooling improved AUC and ACC by 5.3% and 8.9% respectively at the instance level compared to traditional methods, our proposed HAB-MIL achieves superior performance in terms of accuracy, recall, precision and F1, indicating that the effectiveness of the DGA module.

We designed experiments to demonstrate the necessity of using different sequences simultaneously in the model. As shown in Tables 5 and 6, removing the T1w or T2w sequence from the input of the model results in a decrease in IDH prediction performance. Figure 4 shows the ROC curves of different sequences in the TCIA dataset. This indicates that the T1w and T2w sequences are crucial for the IDH prediction task, as they guide the network to focus more on the tumor region and extract features that are highly relevant for gliomas.

To validate the proposed HAB-MIL algorithm for IDH statusclassification more effectively, we compared our proposed HAB-MIL algorithm with eight state-of-the-art IDH prediction methods, including Radiomics approaches (34–36), MDL (37), MultiGeneNet (38), FAD (39), SGPNet (40), PS-Net (41), MFEF Net (24), MTTU Net (42), MTS-UNET (43) and GLISP (44). The results of the different methods are shown in the Table 7. By comparing the proposed method with other IDH prediction networks, the following results can be observed.

He, Bumes, and Kandalgaonkar et al. applied radiomics-based approaches in their studies. In terms of data acquisition, Bumes utilized magnetic resonance spectroscopy (MRS), while He employed contrast-enhanced T1-weighted imaging (T1C). Although multimodal data were incorporated, insufficient feature standardization or inter-modality registration may have hindered the effective integration of information, resulting in amplified noise and reduced model performance. Moreover, conventional radiomics methods are limited to extracting shallow, handcrafted features, which may fail to capture the complex pathological heterogeneity underlying gliomas.

Firstly, GLISP model for patch-level prediction employs a lightweight CNN architecture. While this design ensures computational efficiency, it may fall short in capturing deeper and more complex pathological features compared to more advanced models. Secondly, the prediction accuracy of MultiGeneNet is higher than that of FAD. This is because MultiGeneNet is designed to simultaneously predict multiple key genetic mutations. It employs a shared feature extraction backbone with separate classification branches for each task, enabling effective feature sharing while preserving task-specific distinctions—ultimately improving overall performance. However, the model processes patch samples from whole-slide images without explicitly modeling intra-tumoral heterogeneity, which may result in the loss of important regional variations within the tumor.

Then MTS-UNET architecture lacks an explicit positional encoding mechanism, which may hinder its ability to capture spatial context—such as the relative positioning of lesions—particularly in scenarios involving high tumor heterogeneity. The poor performance of MTDL may be due to the loss of information on intra-tumoral heterogeneity and the variation in tumor size between patients. The model searches for the largest tumor bounding box and then crops the input image to a fixed large size without incorporating the tumor mask information. This approach can introduce irrelevant background information into the model, particularly when the tumor size is small.

Although the MFEF net model incorporates advanced modules such as segmentation-guided feature extraction, asymmetric amplification, and dual-attention feature fusion, the asymmetric amplification module may amplify not only pathological features, but also noise. Additionally, the feature spaces of the T2w and Flair sequences may differ significantly. Calculating these differences directly can introduce inconsistencies, affecting the stability of feature representation and prediction performance, resulting in suboptimal outcomes. MTTU Net achieved good results using a CNN transformer encoder, but it employs an uncertainty-aware pseudo-label selection method to generate and select pseudo-label. If the initial model generates a significant number of incorrect pseudo-labels, these errors may propagate and amplify with each iteration, causing the model to gradually deviate from the correct target and fall into a negative feedback loop.

Hence, the proposed HAB-MIL model is designed to address the limitations of approaches by explicitly accounting for the high heterogeneity of gliomas. It integrates positional encoding during the feature extraction phase and leverages the DGA module to emphasize the significance of key instances. As a result, it achieves superior performance in IDH status prediction compared to all other evaluated methods.

Figure 5 illustrates how our proposed HAB-MIL model localizes discriminative regions for the prediction of IDH status. In particular, the three columns of images are shown for each subject: (1) Original image of different patients, (2) key patches identified by HAB-MIL, and (3) Grad-CAM visualizations produced by HAB-MIL. According to Figure 5, the Key patches in wild-type IDH tumors are more widely distributed, encompassing both the tumor core and surrounding regions. In contrast, key patches in mutant-type IDH tumors are predominantly confined to the tumor core or adjacent areas, indicating a more localized distribution and reduced invasiveness.

The boundaries of wild-type IDH tumors appear relatively indistinct, and some tumor regions show low contrast to surrounding tissues, suggesting a tendency to invasive growth. In comparison, mutant-type IDH tumors show well-defined boundaries, concentrated lesions, and more prominent high signals in the tumor core, indicating a denser internal structure. These observations highlight significant differences in morphological characteristics, invasiveness, and growth patterns between the two types of tumors, offering valuable imaging-based insights to guide treatment decisions and prognosis assessment.