Seeking an optimal approach for Computer-aided Diagnosis of Pulmonary Embolism

Pre-processing.

Similar to the first-place solution for the RSNA Pulmonary Embolism Detection Challenge (Colak et al., 2021b), lung localization and windowing have been used as pre-processing steps. Lung localization removes irrelevant tissues and keeps the region of interest in the slices, whereas windowing highlights the pixel intensities within the range of [100, 700 HU]. Also, the slices are resized to 576 × 576 pixels. Fig. 1 illustrates these pre-processing steps in detail. We consider three adjacent slices from an exam as the 3-channel input of the model.

Pre-processing.

We take advantage of the provided ground truth masks to generate slice-level annotations. Similarly, we apply the same pre-processing steps to the dataset as we did for the RSNA PE dataset. We leverage the best CNN model trained on the RSNA PE dataset to assess the model’s performance using the CAD-PE challenge dataset (Section 5.1.1).

Pre-processing.

The FUMPE dataset includes segmentation masks that highlight the regions of PE. We investigate the dataset to identify which slices have mask values, and consider slices with mask values to be PE positive. We apply the same pre-processing procedure to the FUMPE dataset as we did for the RSNA PE dataset. We use the optimal CNN model trained on the RSNA PE dataset to evaluate the model’s performance on the FUMPE dataset (Section 5.1.1).

Pre-processing.

The approach developed in Liang and Bi (2007), Bi and Liang (2007) has been utilized for generating PE candidates based on heuristic lung segmentation and the tobogganing algorithm (Fairfield, 1990). The lung appears hypoattenuating or darker than its surroundings in chest CTPA scans. Therefore, to separate the lungs from the remainder of the scan, the voxel intensity values are clipped using a −400 HU threshold, resulting in a binary volume with the lungs and other dark regions appearing hyperattenuating or white. Then, a closing operation is performed to fill all dark gaps in the white region. A 3D connected component analysis eliminates non-lung areas, removing components with small volumes or a considerable length ratio between the major and minor axes. The segmentation of the lungs is intended to reduce computational time and the frequency of false positives for the tobogganing algorithm. As PEs exclusively appear within pulmonary arteries, exploring PE candidates outside the lungs is not required. The tobogganing algorithm is then applied specifically to the lung area, generating the PE candidate coordinates used to crop sub-volumes from the CTPA scan. The candidate generation process is detailed in Liang and Bi (2007), Bi and Liang (2007). After candidate generation, each candidate is labeled as “PE” or “non-PE” based on the clot-based ground truth masks. Tajbakhsh et al. (2015) developed a 3D vessel-oriented image presentation: A principal component analysis was performed to determine the vessel axis and the two orthogonal directions. By rotating the orthogonal direction, a number of cross-sectional and longitudinal image planes were obtained. Finally, a 3-channel image representation was created for each PE candidate by selecting the middle slices from each axis. This 3-channel image representation of each PE candidate is used to form the 2.5D orthogonal PE data. The 2D slice-based PE data was formed by copying the middle slice of the z-axis 3 times. Last, each of the candidates themselves from the patients was used as 3D volume-based PE data. Fig. 2 shows both standard PE representation and vessel-oriented PE representation appearances. We should note that the 2.5D image presentation is equivalent to that reported by Tajbakhsh et al. (2015), while the 2D image presentation is included in this paper as a baseline, and the 3D vessel-oriented image presentation has not been used in earlier studies.

4.1.1. Fine-tuning fully-supervised pre-trained models

In transfer learning, a model pre-trained on a different source task is fine-tuned on the target dataset (Zhou et al., 2017a). In this set of experiments, models pre-trained on ImageNet under a fully-supervised setting are fine-tuned for PE slice-level classification using the RSNA PE dataset. For this purpose, we have extensively examined 12 different CNN architectures (Fig. 4), Vision Transformer (ViT), and Swin Transformer. Also, inspired by SeResNext50 and SeResNet50 (Xie et al., 2017), we utilize squeeze and excitation (SE) blocks to the Xception architecture (SeXception).¹ We used only one epoch to train the models. The learning rate, batch size, and optimizer was 0.0004, 20, and Adam, respectively. We used four V100 GPUs to train and test the models. We have also explored the usefulness of ViT, in which the slices are reshaped into a sequence of patches. Upscaling the slice for a given patch size will effectively increase the number of patches, thereby enlarging the size of the training dataset. Similarly, the number of patches increases with a decrease in the patch size. Hence, to explore these two characteristics, we experimented with 32 × 32 and 16 × 16 patches, respectively (Devlin et al., 2018), and also with slices of different sizes. Here, we explored both Vision Transformer (ViT) and Swin Transformer pre-trained on ImageNet1k and ImageNet21k.

4.1.2. Fine-tuning self-supervised pre-trained models

Self-supervised learning eliminates the need for explicit annotations by training the model on a pretext task. The pre-trained model can then be fine-tuned on downstream tasks such as classification and segmentation, as demonstrated in Hu et al. (2021). One example of a pretext task is reconstructing the original image from a distorted version using strong augmentations. When it comes to transfer learning, pre-trained SSL models can be used to initialize the model’s weights, instead of using randomly initialized weights or weights from supervised ImageNet models. In our study, we collected 19 publicly available pre-trained SSL models and fine-tuned them on the RSNA PE dataset for the task of PE slice-level classification (as shown in Fig. 8). All of the models used ResNet50 (He et al., 2016) as the backbone architecture. The aim of this experiment is to compare the performance of models initialized with self-supervised learning against those initialized with fully-supervised learning, in order to understand the benefits of using SSL models in transfer learning.

4.2.1. Sequence model learning

As for sequence model learning, we employed the bidirectional Gated Recurrent Unit (BiGRU). BiGRU is a sequence model consisting of two GRUs (Cho et al., 2014). One GRU takes an input sequence in a forward direction while the other one proceeds in a backward direction. The first-place solution in the RSNA challenge used a single bidirectional Gated Recurrent Unit (BiGRU) to generate an output sequence. The output sequence goes through a max pooling and an attention weighted average pooling and then is concatenated as an exam-level embedding. The exam-level embedding undergoes 10 separate classification layers to predict nine exam-level labels and one slice-level label. The hidden size of the BiGRU is 512, and the architecture is trained for 25 epochs with a batch size of 64. The current implementation of the classification layer in the BiGRU model, which predicts slice-level labels, is flawed. As the number of slices per exam is inconsistent, the design uses linear interpolation on the slice-level ground truth to achieve a fixed size of 192, but in doing so, the expert annotation is compromised. A more effective approach would be to adopt a framework that is capable of handling variable numbers of slices and incorporates expert annotation to ensure more accurate predictions.

4.2.2. Embedding-based Vision Transformer (E-ViT)

We modified Vision Transformer (ViT) (Dosovitskiy et al., 2020) for the exam-level diagnosis as illustrated in Fig. 3. First, E-ViT takes a sequence of slice-level embeddings rather than image patches as input. The slice-level embeddings are extracted from the task of slice-level classification (Section 4.1), so they are more discriminative and informative than raw CT slices. Our E-ViT framework works with a varying number of slices, unlike the first-place solution which uses a fixed number of slices. Second, we use 10 class tokens (nine for each exam label and one for slice label) to address the multi-label classification task. Having the slice label alongside exam labels enforces the consistency between exam-level and slice-level predictions. Third, E-ViT discards the position embedding for each CT slice because neighboring slices in a CT scan appear similar; therefore, retaining the position embedding can lead to exam-level performance degradation, as evidenced in Table 6. In summary, E-ViT integrates class embedding with exam-level embedding to fully exploit the feature representations extracted from individual sequences with varying numbers of slices.

We apply the same pre-processing to the embeddings extracted from the slice-level classification stage (presented in Section 4.2.1). We then incorporate a linear layer to reduce the slice-level embedding dimension from V × 6144 to V × 768 as the input of the Transformer encoder. Here, V represents the varying number of slices per exam as the number of slices is not consistent across different exams. Through our investigation of the RSNA PE dataset, we found that the upper bound of V can be set to 512, as it is sufficient for the detection of PE within the lung area. When working with varying numbers of slices per exam, it is important to consider how best to utilize the available data in order to ensure accurate predictions. In our approach, If the number of slices for a particular exam is greater than 512, we use the attention masking mechanism from the BERT implementation (Devlin et al., 2018) to ensure that only the relevant slices are used in the model. This allows us to focus on the most relevant information and improve the efficiency of the model. On the other hand, if the number of slices for a particular exam is above 512, we assume that most likely the slices contain regions other than the lungs. Therefore, we simply use the 512 slices belonging to the lung area to ensure that the model is only processing relevant information. This approach allows us to effectively handle varying numbers of slices and optimize the use of the available data.

The Transformer encoder uses a multi-head self-attention mechanism to output contextualized slice-level embedding (sized V × 768) and class embedding (sized 10 × 768). We concatenate the contextualized slice-level embedding which underwent attention weighted average pooling and max pooling together. The concatenated exam-level embedding is then fed to 10 separate classification layers to predict the nine exam-level labels and one slice-level label. Similarly, each class embedding is utilized to predict the corresponding label through 10 different classification layers. The class embedding and exam-level embedding are integrated using the formula Avg(CE, AMP(EE)). Here, CE represents class embedding, and AMP (EE) represents the exam-level embedding (EE) which is the concatenation output of attention weighted average pooling and max pooling for contextualized slice-level embedding from the Transformer encoder. We name this architecture Embedding-based Vision Transformer (E-ViT). Also, this architecture is trained with the mean binary cross entropy loss from exam-level embedding and class embedding predictions. During inference, the mean probability score from these two branches is utilized for making the final exam-level diagnosis. In regards to the initial set of hyperparameters, we have utilized the original Vision Transformer (Dosovitskiy et al., 2020), specifically ViT-B. However, to further optimize the performance of our proposed E-ViT model, we conducted several experiments to find the best combination of hyperparameters such as learning rate, optimizer, batch size, and total number of epochs. According to our experiments, the best combination of hyperparameters for the E-ViT model are reported in Table 1.

5.1.1. Fine-tuning pre-trained models outperforms training from scratch

Fig. 4 shows a significant performance gain for every pre-trained model compared with random initialization. There is also a moderate positive correlation of 0.5914 between ImageNet and PE classification performance across different architectures (Fig. 5). The correlation indicates that the useful weights learned from ImageNet can be successfully transferred to the PE classification task, despite the modality difference between photographic images (ImageNet) and CTPA scans (RSNA PE). To demonstrate the interpretability of the trained model, we use GradCam++(Gildenblat and contributors, 2021) to visualize the attention map of the SeXception architecture, which performed the best as a standalone model. As shown in Fig. 6, the attention map effectively highlights the potential PE location in the slice. Additionally, to assess the generalizability of the SeXception model trained on the RSNA PE dataset, we directly test the model on two additional datasets: CAD-PE (González et al., 2020) and FUMPE (Masoudi et al., 2018). The results are summarized in Table 2. As shown in Fig. 7, the GradCam++ highlights the PE region. It is noteworthy that the model was only trained on the RSNA PE dataset and had not seen the two additional datasets prior to evaluation. We evaluated the two additional datasets for slice-level classification to determine whether a slice has PE or not. Here, the model had not seen the segmentation ground truth for these datasets either. Despite this, the model not only accurately classified slices with PE but also efficiently localized them. The green highlighted area in the figure represents the ground truths of the given segmentation mask, whereas the yellow highlighted area represents the predicted PE region. The experiments demonstrate not only the feasibility of using deep learning for diagnosing challenging radiologic findings such as PE on CTPA, but also the ability to accommodate data from external institutions employing different CT scanners and imaging protocols. Moreover, the superior performance of fine-tuning pre-trained models highlights the importance of transfer learning in medical image analysis and could have implications for improving the diagnostic accuracy and efficiency of PE detection in clinical practice. The SeXception model, trained on the RSNA PE dataset, demonstrates strong performance in PE slice-level classification on the CAD-PE Challenge and FUMPE datasets. According to Table 2, the SeXception achieves a sensitivity of 87.57% and a specificity of 87.60% on the CAD-PE Challenge dataset, and a sensitivity of 83.85% and a specificity of 83.95% on the FUMPE dataset. We also combine the slice-level classification result on the RSNA PE dataset by taking the average from Xception & SeXception, SeResNext50 & SeXception, and SeResNext50 & Xception & SeXception Table 4. The ensemble of SeResNext50 & Xception & SeXception achieves the best AUC performance of 96.76 ± 0.05 gaining 0.62% compared with the first-place solution.

5.1.2. Self-supervised pre-training can surpass fully-supervised pre-training

As summarized in Fig. 8, SeLa-v2 (Asano et al., 2020) and DeepCluster-v2 (Caron et al., 2018) have achieved the best AUC of 95.68 ± 0.05 and 95.68 ± 0.06, followed by Barlow Twins (Zbontar et al., 2021). Seven of 19 self-supervised ImageNet models performed better than supervised pre-trained ResNet50 (Fig. 8). Six self-supervised methods underperform fully-supervised pre-trained ResNet50 but outperform learning from scratch. Finally, six self-supervised methods underperform ResNet50 learning from scratch. These results also highlight the importance of the pretext tasks for self-supervised learning, and the performance can vary significantly with these tasks for the same backbone. In the context of clinical practice, the self-supervised pre-training provides a good initialization for fine-tuning the downstream task of PE slice-level classification. However, an unfavorable pretext task leads to inferior performance on the downstream task compared to scratch training. We believe that a good pretext task based on large-scale PE detection will provide a stronger initialization for robust and efficient slice-level PE classification. On the other hand, SSL can be useful in the clinical practice of PE slice-level classification due to its ability to learn from unlabeled data. In a clinical setting, obtaining annotated data for PE classification can be challenging and time-consuming, and SSL can leverage the large amounts of available unlabeled data to learn a representation that can be used for classification tasks.

The details and comparison of all these self-supervised methods are provided in the appendix (see Table A.2).

5.1.3. Transformer-based models demonstrate comparable performance with slower convergence speed

Table 3 provides compelling evidence that the Swin Transformer (Swin-B), which was pre-trained using the SimMIM self-supervised approach, outperforms other transformer-based models, attaining an impressive AUC score of 96.46, even with a larger image size of 448 × 448. Notably, Swin-T, pre-trained with ImageNet1k, achieves a commendable AUC score of 96.30, coming remarkably close to the performance of the standalone CNN architecture (SeXception). However, it is important to note that fine-tuning transformer-based models required substantially more time and a greater number of training epochs to reach convergence. Also, the performance of the transformer-based models varies significantly with the training set size. As the transformer-based models are trained using the image patches, two approaches are followed to expand the training set:

1. The patch size is decreased for a given image size, effectively increasing the total number of patches.

2. The image size is increased for a given patch size, thus enhancing the total patch count.

The results demonstrate the crucial role of initialization as pre-training for transformer-based models, highlighting its significance (Ma et al., 2022). In contrast to CNNs, the number of parameters was large for transformer-based models and it took more than 16 h to finish one epoch of training. Although we have a huge number of slices, CNNs took only about 3 h to complete one epoch of training. Details are provided in the appendix (see Table A.3).

5.1.4. Squeeze and excitation (SE) block further enhances CNN performance

Despite fewer parameters compared with many other architectures, SeXception provides an optimal average AUC of 96.34 ± 0.09. Thus, the SE blocks are parameter-efficient and have led to performance improvements for a variety of CNN architectures, such as ResNet50, ResNet101, ResNext50, ResNext101, and Xception (Fig. 9). This observation is consistent with (Hu et al., 2018a), which infer SE blocks capture feature dependencies in the channels to improve the performance. Here, SE blocks in a CNN enhance PE detection by finding relationships between consecutive CT slices, thereby improving the model’s ability to identify and classify PEs with higher accuracy and efficiency. Additionally, the SE blocks facilitate learning more discriminative features and improve the handling of large variability in medical images, resulting in enhanced accuracy and robustness in PE slice-level classification for CTPA data. Consequently, the utilization of SE blocks is deemed a valuable tool in clinical practice.

5.2.1. Sequence models benefit from more comprehensive slice-level embeddings

Table 4 summarizes the results of the sequential model learning (explained in Section 4.2.1) for exam-level diagnosis. Although the combination of SeResNext50, Xception and SeXception performs optimally for slice-level classification (see Table 4), this is not the case for exam-level diagnosis. As a standalone CNN backbone, Xception achieves the best mean AUC score across nine exam-level labels gaining 1.05% compared with the previous state of the art (SeResNext50). We also combine the exam-level diagnosis result by taking the average from Xception & SeXception, SeResNext50 & SeXception, and SeResNext50 & Xception & SeXception. We observe that Xception & SeXception achieve a significant gain of 1.21% (p = 5.94E−13) over the state of the art. Subsequently, we harnessed more extensive slice-level embeddings derived from Swin-B and Swin-T to enhance the exam-level diagnosis (Table 5). So far, Swin-T, acting as the backbone, has established itself as the leading performer for exam-level PE diagnosis, demonstrating an impressive AUC of 90.11 and achieving a notable improvement of 0.83 (p = 6.81E−14) compared to the ensemble of Xception and SeXception.

5.2.2. Our E-ViT further exceeds sequence models

Table 5 demonstrates the mean AUC for all the exam-level predictions with the sequence model and our proposed Transformer-based model. In our evaluation, utilizing only class embedding for exam-level diagnosis results in performance comparable to the sequence model. Conversely, utilizing only the exam-level embedding from the Transformer encoder leads to superior performance, surpassing both the sequence model and E-ViT with class embedding alone. The integration of class embedding and exam-level embedding in our E-ViT framework results in even further improved performance for exam-level diagnosis. These observations are consistent across slice-level embeddings extracted from three different backbones (i.e., SeResNext50, Xception & SeXception). When the slice-level embeddings are extracted from SeResNext50 and used as inputs, our E-ViT outperforms sequence model by an AUC gain of 1.01%. We also analyze the performance achieved by combining two or three architectures together by simply taking the mean prediction of each exam-level across slice-level embeddings extracted from different backbones. The best AUC performance (89.63) is achieved from the ensemble of Xception and SeResNext50, gaining 1.56% over the first-place solution with a statistical significance of 2.19E-11 when using CNN architecture as the backbone. We further utilized slice-level embeddings from Swin-B and Swin-T for our E-ViT for the task of exam-level PE diagnosis. Notably, our findings revealed that the pairing of Swin-T as the backbone with our E-ViT yielded the most impressive performance, achieving an AUC of 90.29 for exam-level diagnosis. While the combination of Swin-T and BiGRU exhibited similar results, our E-ViT outperformed it significantly (p = 8.61E−4). In clinical practice, the number of slices in a CT scan can vary greatly between patients, making it challenging to use traditional models that require a fixed number of inputs. Our proposed Transformer-based model, E-ViT, is designed to tackle this issue by adapting to varying numbers of slices without losing any expert annotations. This is a crucial aspect of its use in clinical practice, as it allows for a more efficient and accurate diagnosis of PE. In the following, we present ablation studies on our E-ViT.

5.3.1. 3D data offer higher performance than 2D and 2.5D data

While it may appear that using 3D models to handle 3D volumetric data is a reasonable choice, it comes with a high computing cost and a risk for overfitting (Zhou et al., 2021b). As a result, various alternative methodologies for reformatting 3D applications into 2D solutions have been presented. Regular 2D inputs were constructed by extracting neighboring axial slices (referred to as 2D slice-based input) by Ben-Cohen et al. (2016) and Sun et al. (2017). These 2D reformatted solutions create large amounts of data and take advantage of 2D models pre-trained on ImageNet. However, 2D solutions unavoidably compromise the rich spatial information in 3D volumetric data and the large capacity of 3D models. Alternatively, a more advanced technique, as detailed by Prasoon et al. (2013) and Roth et al. (2014, 2015), is to extract axial, coronal, and sagittal slices from volumetric data (referred to as 2.5D orthogonal input).

Table 8 compares 2D slice-based, 2.5D orthogonal, and 3D volume-based approaches as different image dimensions for PE false positive reduction. According to our analyses, 3D volume-based data consistently outperforms 2D and 2.5D data because it contains significantly more information. When the model is trained from scratch, 3D volume-based data outperforms 2.5D orthogonal data in both VOIR and non-VOIR scenarios (78.81% vs. 80.07% and 86.51% vs. 91.35%). Here, the 3D volume-based data significantly outperforms the 2D slice-based and 2.5D orthogonal approaches, with p-values of 7.94E−13 for non-VOIR and 2.19E−10 for VOIR scenarios, respectively. These results indicate that the use of 3D data in CTPA can lead to a reduction in false positive results compared to 2D or 2.5D data. This is because 3D data provides a more comprehensive and detailed view of the blood vessels in the lungs, which can aid in the accurate identification and differentiation of true positives from false positives. Overall, the use of 3D data in CTPA can improve the accuracy and reliability of the results, leading to better patient care in the clinical context.

5.3.2. VOIR is more informative than the standard image representation, boosting performance across image dimensions

Developing a computer-aided diagnosis (CAD) system that can accurately detect pulmonary embolism (PE) while maintaining a clinically acceptable level of false positives is highly desirable (Tajbakhsh et al., 2016). CTPA is the primary imaging modality used to diagnose PE, and appropriate pre-processing of these images is necessary to ensure accurate detection and reduce false positives. One approach to improve the performance of the model is through the use of vessel-oriented image representation (VOIR) (Tajbakhsh et al., 2019), which promotes consistency in the data and has been shown to be effective for PE detection. However, previous implementations of VOIR were limited to a 2D representation and did not fully exploit the 3D information available in CTPA images. In this study, we extend VOIR to a 3D form (Section 4.3) and evaluate its performance under various settings to improve the CAD system’s ability to accurately detect PE and reduce false positives for better clinical outcomes.

Table 8 summarizes the results with VOIR and non-VOIR using 2D, 2.5D, and 3D input data as different image dimensions. The AUC is higher with VOIR data in all image dimensions (2D, 2.5D, and 3D). As discussed in Section 5.3.1, although 2.5D data is expected to be more valuable than 2D data (Zhou, 2021), our results show that 2D data with VOIR significantly outperforms 2.5D data non-VOIR by an AUC gain of 7.21%. Similarly, 2D slice-based VOIR data outperforms 3D volume-based non-VOIR data (5.95% gain), even though 3D data contains significantly more information. These results show that vessel-oriented image representation is more informative than the standard image representation for reducing PE false positives.

5.3.3. Same domain transfer learning with self-supervised pre-training enhances performance across image representations and dimensions

One of the most commonly utilized paradigms in deep learning for medical image analysis is transfer learning from photographic images to medical images. Medical images, such as those used in radiology and diagnostic imaging, are distinct from photographic images in that they are often monochromatic and feature consistent anatomical structures. These characteristics can make medical images more challenging to interpret and analyze (Hosseinzadeh Taher et al., 2021). Therefore, same-domain transfer learning should be applied since a smaller domain gap makes the learned image representation more effective for target tasks (Zhou et al., 2021b). Models Genesis (Zhou et al., 2019), a self-supervised method, were developed to mitigate the domain gap. Specifically, Models Genesis were pre-trained on LUNA16 (Setio et al., 2016), a chest CT dataset that shares similar body parts and imaging modality as our PE-CAD dataset. We use Models Genesis with ResNet18 as the backbone for the task of PE false positive reductions. We also evaluate the recently published 3D self-supervised approaches (Guo et al., 2022), which were pre-trained on the LUNA16 dataset, to perform the same task. We anticipate improved performance in image representation (non-VOIR vs. VOIR) and dimensions (2D vs. 2.5D vs. 3D) compared to supervised pre-trained models due to their self-supervised nature.

According to Table 8, weights initialized from Models Genesis consistently outperform the ImageNet counterpart. For 2D slice-based non-VOIR data, Models Genesis initialization achieves an AUC of 72.11% ± 1.45 compared to 63.29% ± 1.54 with ImageNet initialization, achieving a significant gain of 8.82% (p-value = 2.63E−11). Similarly, for 2.5D orthogonal non-VOIR inputs, the gain is 2.98% with a p-value of 1.34E−9. A similar performance is observed for VOIR data, with a gain of 0.93% and 1.79% for 2D and 2.5D input data, respectively. Hence, self-supervised pre-training with Models Genesis enhances performance across image representations and dimensions. The results demonstrate the advantages of using the same domain transfer learning in medical image analysis. This approach involves utilizing models trained on similar medical image data to perform a new task. For instance, we found that transfer learning within the same domain is advantageous in medical image analysis as it allows the model to utilize its previously learned features to efficiently adapt to a new task (Zhou et al., 2019). As a result, same domain transfer learning can significantly enhance the accuracy and efficiency of medical image analysis in the clinical context, ultimately leading to better patient care and outcomes. Furthermore, we demonstrate our results by conducting an evaluation of the latest 3D self-supervised pre-trained models on the 3D VOIR dataset. According to Table 9, the TransVW model with the pre-training strategy of ((D) + R) + A performs best in reducing false positives on 3D VOIR data, achieving an AUC of 93.41%. Here, D represents the discriminative encoder, R is the restorative decoder, and A the adversary encoder.