Improving deep neural network generalization and robustness to background bias via layer-wise relevance propagation optimization

TL;DR

The paper tackles background bias and shortcut learning in deep classifiers by introducing BRM, an optimization of differentiable LRP heatmaps during ISNet training. ISNet reduces reliance on background features without increasing run-time cost, and is compatible with virtually any backbone. Across synthetic bias, COVID-19, and TB datasets, ISNet demonstrates superior out-of-distribution generalization and foreground-focused attention, outperforming eight state-of-the-art baselines. The approach is theoretically grounded in Deep Taylor-based LRP, showing denoising benefits and more stable optimization, and is validated with heatmaps that align with foreground regions and radiologist annotations. Overall, BRM via LRP optimization offers an efficient, versatile method to mitigate background bias in clinical and natural image classification tasks, improving real-world reliability.

Abstract

Features in images' backgrounds can spuriously correlate with the images' classes, representing background bias. They can influence the classifier's decisions, causing shortcut learning (Clever Hans effect). The phenomenon generates deep neural networks (DNNs) that perform well on standard evaluation datasets but generalize poorly to real-world data. Layer-wise Relevance Propagation (LRP) explains DNNs' decisions. Here, we show that the optimization of LRP heatmaps can minimize the background bias influence on deep classifiers, hindering shortcut learning. By not increasing run-time computational cost, the approach is light and fast. Furthermore, it applies to virtually any classification architecture. After injecting synthetic bias in images' backgrounds, we compared our approach (dubbed ISNet) to eight state-of-the-art DNNs, quantitatively demonstrating its superior robustness to background bias. Mixed datasets are common for COVID-19 and tuberculosis classification with chest X-rays, fostering background bias. By focusing on the lungs, the ISNet reduced shortcut learning. Thus, its generalization performance on external (out-of-distribution) test databases significantly surpassed all implemented benchmark models.

Figures (11)

• Figure 1: Heatmaps (Layer-wise Relevance Propagation/LRP for convolutional networks and attention rollout for Vision Transformer) for positive COVID-19 and Pneumonia X-rays and photographs, extracted from the synthetically biased test datasets (biased test). Last row displays classifier trained without the synthetic bias (and analyzing images without the bias), for reference. The Image's true class is stated above the figures, the DNN that produced the heatmap is identified on the left. The triangle (background bias) indicates the classes COVID-19, smiling or Pug. The circle pneumonia, high cheekbones, and Tibetan Mastiff. The square rosy cheeks and Pekingese. Red colors in the LRP maps indicate areas the DNN associated to the image's true class, while blue colors are areas that reduced the network confidence for the class. For attention rollout, red shows the DNN attention. White represents areas with little influence over the classifiers. In the heatmaps, focus on the images' foregrounds (dogs, faces or lungs) is desirable. For privacy, the face picture was substituted by a representation of the face (gray) and bias (white) locations, but classifiers received the real picture
• Figure 2: Heatmaps (Layer-wise Relevance Propagation/LRP for convolutional networks and attention rollout for Vision Transformer) for positive COVID-19, Pneumonia and tuberculosis. The Image's true class is stated above the figures, the DNN that produced the heatmap is identified on the left. For LRP, red colors indicate areas that the DNN associated to the true class, blue colors are areas that decreased the network confidence for the class. For attention rollout, red indicates the DNN attention. White represents areas with little influence over the classifiers. Attention outside of the lungs (foreground) is undesirable
• Figure 3: LRP (Layer-wise Relevance Propagation) heatmaps and Grad-CAM (Gradient-weighted Class Activation Mapping) for X-rays and photographs. Image classes presented on the left, heatmap type and classifier on the top. In the Grad-CAMs, the redder the area, the more attention was paid to it. Red colors in the LRP maps indicate areas the DNN associated to the ground truth classes. Blue colors in LRP heatmaps are areas that reduced the network confidence for the classes. White indicates little DNN attention. a Images from the COVID-19 and Tuberculosis detection datasets, without synthetic bias, and heatmaps from classifiers trained on data without synthetic bias. b Images extracted from the artificially biased test datasets, and classifiers trained on synthetically biased data. The triangle (background bias) indicates the presence of the classes COVID-19, smiling or Pug. The circle indicates pneumonia, high cheekbones and Tibetan Mastiff. For privacy, the face picture was substituted by a representation of the face (gray) and bias (white) locations, but classifiers received the real picture
• Figure 4: X-rays and their LRP (Layer-wise Relevance Propagation) heatmaps (blue and red), superposed over the image annotated by the radiologist (magenta lines). Red colors indicate areas that the DNN associated to the true class (indicated over the X-rays), blue colors are areas that decreased the network confidence for the class
• Figure 5: Representation of an ISNet for a classifier containing 3 layers. The classifier is on the left, and the corresponding LRP (Layer-wise Relevance Propagation) block on the right. L1 and L2 represent convolutional layers, L3 represents a fully-connected layer. $\mathbf{x_{i}}$ indicates the input of classifier layer Li. With LRPi being the layer in the LRP block responsible for performing the relevance propagation through Li, its output, $\mathbf{R_{i-1}}$, is the relevance at the input of layer Li, and at the output of layer Li-1. $\mathbf{y}$ is the classifier's output, and $\mathbf{y'}$ is the output for one of the classes (with all other logits set to zero), which will be used to generate the heatmap associated with the class. $\mathbf{Input}$ is the classifier's input and $\mathbf{Heatmap}$ its LRP heatmap. Convolutional layers and their LRP counterparts are displayed in orange, while fully-connected layers and their LRP counterparts are in yellow