Experiments

In this section, we present our benchmarking of state-of-the-art saliency methods using the proposed datasets. We begin by describing the detailed experimental settings in Section 5.1, followed by a comprehensive evaluation and analysis of the results in Section 5.2.

Experimental Settings

Baselines

We benchmark six common saliency methods: GradCAM [48], GradCAM++ [10], Integrated Gradients [53], InputXGradient (IxG) [49], Occlusion [62], and RISE [44] in our eight proposed datasets. The experiments were conducted using two different CNN architectures: ResNet-18 [23] and VGG-19 [51]. We also benchmark the attention mechanism as a saliency explanation method of ViT with ViT-base-patch16–224 architecture (ViT-B/16) [15].

Evaluation metrics

We evaluate the performance of the saliency methods using three metrics: mIoU and PG for alignment-based evaluation and iAUC for faithfulness-based evaluation. For all metrics, the overall results are obtained as the average across all the samples in the test set.

Implementation Details

In each experiment, we allocated the data as follows: 70% for training, 15% for validation, and the remaining 15% for testing. Both pretrained ResNet-18 and VGG-19 models were trained for 30 epochs using the Adam optimizer with a learning rate of 0.001, and the best checkpoint was selected based on performance on the validation set. The implementation of the saliency methods followed the original settings described in the respective papers. Our experiments are implemented based on PyTorch and performed on an NVIDIA A10G GPU.

Results and Analysis

In this section, we present a detailed analysis of the results obtained from our benchmarking experiments, as summarized in Table 2.

Results and Analysis for ResNet-18.

From the results for ResNet18, we observe a generally strong performance across all datasets. Notably, for the Gender-XAI dataset, the best performance in terms of mIoU is achieved by the Integrated Gradients (0.3441), followed closely by RISE (0.3416). The Occlusion method shows relatively lower performance with a mIoU of 0.1301, indicating that its effectiveness in identifying the regions relevant for model predictions is lower than the other methods. For the Environment-XAI dataset, RISE also outperforms the other methods, yielding a mIoU of 0.5428. This is consistent with the findings from GradCAM++, which ranks second in performance. In contrast, Integrated Gradients shows slightly weaker performance in this case, especially in terms of PG (0.2810), which suggests that its ability to highlight important regions of an image could be improved. The Action-XAI and Object-XAI datasets show that RISE consistently provides robust performance in mIoU (0.4406 and 0.5430, respectively), followed by GradCAM and GradCAM++. Disease-XAI and Cancer-XAI datasets display a notable performance gap, where GradCAM tends to perform better in identifying salient regions compared to the other methods, suggesting that GradCAM’s gradient-based approach works better for these types of medical images.

Results and Analysis for VGG-19.

The results for VGG-19, while showing similar trends, reveal that VGG-19 generally performs slightly worse than ResNet-18 across most datasets, which is consistent with the performance gap typically observed between these two architectures. The Gender-XAI dataset shows that RISE remains the top performer in terms of mIoU (0.2961), though GradCAM++ (0.2859) performs comparably, indicating that the additional information from the gradients in GradCAM++ may be more useful in this case. For Environment-XAI, RISE again excels (mIoU of 0.4792), and GradCAM++ performs slightly worse than ResNet-18 at 0.3160. However, RISE’s consistently strong performance across datasets reinforces its utility in identifying salient regions for model predictions. The Disease-XAI and Cancer-XAI datasets present challenges for the saliency methods, with lower mIoU values across the board, particularly for Integrated Gradients (0.1105 for Disease-XAI and 0.0803 for Cancer-XAI), suggesting that the method may struggle with certain types of medical images. In contrast, GradCAM and RISE show more reliable results for these datasets.

Comparative Analysis.

When comparing ResNet-18 and VGG-19, it is evident that ResNet-18 tends to outperform VGG-19 across most datasets, as indicated by higher mIoU, PG, and iAUC values. This is particularly noticeable for the Action-XAI, Object-XAI, and Environment-XAI datasets, where ResNet-18 consistently leads in terms of both alignment and faithfulness metrics. This difference in performance can likely be attributed to the architecture of the models, with ResNet-18 being a deeper network with skip connections that may help it capture more complex features relevant to saliency mapping. Interestingly, the GradCAM++ method shows stronger performance in VGG-19 compared to ResNet-18 on the Gender-XAI dataset, possibly due to the deeper layers in VGG-19 focusing more effectively on the critical regions of images in gender classification tasks. However, RISE and GradCAM remain the most reliable across the board, showing consistent performance on multiple datasets, regardless of the underlying architecture.

Inter-method reliability analysis.

Inter-method reliability analysis examines the consistency of evaluation metrics across different saliency methods, providing insights into whether the metrics yield stable and comparable assessments of explanation quality. We performed this analysis to assess the correlation between mIoU, PG, and iAUC metrics for two base models, ResNet-18 and VGG-19. The results in Table 3 show generally positive correlations across all metrics, with stronger associations observed for ResNet-18. Notably, mIoU and PG have the highest correlation for Integrated Gradients (0.8462), indicating good alignment between the saliency maps and ground-truth annotations. Similarly, mIoU and iAUC are strongly correlated, especially for Integrated Gradients (0.7024). For VGG-19, the correlations are slightly weaker, with GradCAM++ showing the highest PG and iAUC correlation (0.5489). These results suggest that different saliency methods exhibit varying levels of alignment and faithfulness, with some methods (e.g., GradCAM++) being more consistent across metrics than others.

Benchmarking on Vision Transformer.

In this section, we present the benchmarking results of saliency methods using the ViT. Specifically, we employed the ViT-base-patch16–224 architecture [15] (ViT-B/16) and focused on evaluating the attention mechanism as a saliency explanation method. The results, presented in Table 4, include evaluations of mIoU, PG, and iAUC across eight datasets.

As illustrated in Figure 4, generates attention maps based on its transformer architecture, which we directly analyze in this benchmarking. The attention maps produced by ViT are capable of capturing more detailed features of regions of interest compared to traditional saliency methods based on CNNs such as GradCAM and IxG. These maps show a more refined focus on relevant areas, leading to improved performance in the evaluation metrics. This improved performance can be attributed to ViT’s self-attention mechanism, which allows the model to attend to long-range dependencies and contextual information within the image, providing a more holistic view of the important regions compared to methods that rely solely on local pixel-level information.

Observation and Discussion

(1) The performance of saliency methods varies across datasets, influenced by factors such as task complexity, label distribution, and image characteristics. Simpler tasks generally yield better results, while more complex datasets tend to present challenges for some methods. (2) ViT’s attention mechanism emerges as a promising alternative to CNN-based saliency methods, especially for capturing global context and long-range dependencies. However, its robustness and efficiency relative to traditional CNN-based approaches require further investigation. (3) The inter-method reliability analysis highlights varying correlations between evaluation metrics, indicating that saliency methods show different levels of consistency across tasks and models. This emphasizes the importance of careful method selection depending on the specific evaluation goals. (4) The analysis suggests that more complex datasets often lead to reduced faithfulness in saliency methods. This indicates that incorporating additional context or information, such as object-level segmentation, could enhance alignment and faithfulness, particularly in challenging tasks. (5) Overall, while the evaluated methods provide valuable insights, each has its strengths and limitations. Improvements are needed, particularly for more intricate tasks requiring fine-grained explanations.