Adversarial attacks and adversarial robustness in computational pathology

Abstract

Artificial Intelligence (AI) can support diagnostic workflows in oncology by aiding diagnosis and providing biomarkers directly from routine pathology slides. However, AI applications are vulnerable to adversarial attacks. Hence, it is essential to quantify and mitigate this risk before widespread clinical use. Here, we show that convolutional neural networks (CNNs) are highly susceptible to white- and black-box adversarial attacks in clinically relevant weakly-supervised classification tasks. Adversarially robust training and dual batch normalization (DBN) are possible mitigation strategies but require precise knowledge of the type of attack used in the inference. We demonstrate that vision transformers (ViTs) perform equally well compared to CNNs at baseline, but are orders of magnitude more robust to white- and black-box attacks. At a mechanistic level, we show that this is associated with a more robust latent representation of clinically relevant categories in ViTs compared to CNNs. Our results are in line with previous theoretical studies and provide empirical evidence that ViTs are robust learners in computational pathology. This implies that large-scale rollout of AI models in computational pathology should rely on ViTs rather than CNN-based classifiers to provide inherent protection against perturbation of the input data, especially adversarial attacks.

Fully transformer-based biomarker prediction from colorectal cancer histology: a large-scale multicentric study

“…Originating from the field of natural language processing, transformers are a powerful tool to process sequences and leverage the potential of large amounts of data. Also in computer vision, transformers yield a higher accuracy for image classification in non-medical tasks [Dosovitskiy et al, 2020, are more robust to distortions in the input data [Ghaffari Laleh et al, 2022a] and provide more detailed explainability [Chen et al, 2022]. These advantages of transformers compared to CNNs have the potential to translate into more accurate and more generalizable clinical biomarkers, but there is currently no evidence to support this.…”

“…In many non-medical and medical image processing tasks, transformer neural networks have recently been adopted for computer vision tasks [Dosovitskiy et al, 2020, He et al, 2022, replacing CNNs because of their improved performance and robustness [Ghaffari Laleh et al, 2022a]. Originally proposed for sequencing tasks such as natural language processing, transformer networks show impressive capabilities of learning long-range dependencies and contextualizing concepts in long sequences.…”

Seeing the random forest through the decision trees. Supporting learning health systems from histopathology with machine learning models: Challenges and opportunities

“…Although major challenges related to data governance/ management, 135,[175][176][177][178][179][180][181][182][183] ethical/legal, [184][185][186][187] and environmental 188,189 considerations would need to be addressed, these "dynamic multimodal ML models" may become cost-effective in different populations 169,[190][191][192][193] if conceived as the integrative tools needed to provide precision health 22,164,165,168,[194][195][196] and support some of the iterative cycles of knowledge generation and continuous improvement of "learning health systems". 11,[197][198][199][200]…”

A Transformer‐Based microvascular invasion classifier enhances prognostic stratification in HCC following radiofrequency ablation

“…[14][15][16][17] Transformers, initially designed for natural language processing, have recently been adapted for some medical imageprocessing tasks, replacing CNNs due to their superior performance and robustness. 18,19 By incorporating the self-attention mechanism, CI: .667-.778) for evaluating RFS after RFA. Among the three risk factors, MVI was the most powerful predictor for intrahepatic distance recurrence.…”

“…Convolutional neural networks (CNNs) have become the predominant deep learning (DL) approach for computer vision tasks, enabling the automated correlation of features from unstructured image datasets with MVI labels 14–17 . Transformers, initially designed for natural language processing, have recently been adapted for some medical image‐processing tasks, replacing CNNs due to their superior performance and robustness 18,19 . By incorporating the self‐attention mechanism, transformers can effectively capture relationships between image regions with a greater flexibility 20 .…”