Pith. sign in

REVIEW 2 major objections 6 minor 1 cited by

Human-aligned Deep Learning: Explainability, Causality, and Biological Inspiration

T0 review · 2 major / 6 minor · reviewed 2026-08-16 · deepseek-v4-flash

Pith's one-line read The thesis claims that CNN feature co-occurrence, packaged as a 'causality map,' can substitute for causal knowledge in medical image classification, improving accuracy, explainability, and out-of-distribution generalization.

desk verdict A competent, honestly assembled thesis that re-presents already published work; its central 'causality-driven' claim equates co-occurrence with causation and needs a clear fix before the framework's robustness story can be taken seriously. read the letter →

arxiv 2504.13717 v1 pith:NJ7BJ3LP submitted 2025-04-18 cs.CV cs.AIcs.LGeess.IVq-bio.NC

classification cs.CVcs.AIcs.LGeess.IVq-bio.NC
keywords explainableAImedicalimageclassificationfeatureco-occurrencecausaldeeplearningdomaingeneralizationprototypicalpartout-of-distributionrobustnessbiologicallyinspiredvision
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This thesis argues that deep learning can be aligned with human reasoning by treating explainability, causality, and biological vision as one problem, and it stakes its central technical claim on a simple operational move: feature co-occurrence statistics extracted from CNN feature maps are 'weak causal signals' that a classifier can exploit without any causal knowledge about how the images were made. If that move holds, medical image classifiers can become more accurate, more explainable, and more reliable on out-of-distribution data than ordinary attention-based models, which matters because shortcut learning and domain shift are the main reasons medical AI fails in the clinic. The thesis supports the claim across three connected studies: prototypical-part learning for breast mass classification that a radiologist confirms aligns with clinical reasoning; causality-driven CNN modules that improve classification and explanation quality; and the CROCODILE framework that generalizes across medical domains. The biological-vision study, CoCoReco, extends the same co-occurrence idea into a context-aware attention mechanism, and the net contribution is a parameter-light recipe for causal-looking robustness that does not require intervention data.

What carries the argument

The load-bearing object is the 'causality map': a two-dimensional score field derived from feature co-occurrence statistics computed over a CNN's feature maps, which the paper treats as a proxy for causal relevance in a medical image. Around it the thesis builds a 'causality-factors extraction' module that selects and re-weights feature maps, an attention-inspired enhancement scheme, and, in CROCODILE, a mechanism that injects the map as prior knowledge into a dual-branch architecture that disentangles disease features from domain features; in CoCoReco the same co-occurrence principle appears as a Contextual Attention Block that models visual context. The causality map is what carries the argument from association to the claimed causal robustness, so everything the thesis says about explainable, robust, generalizing classifiers depends on this one operationalization.

What would settle it

Train a causality-module classifier on chest X-rays into which a task-irrelevant but consistently co-occurring artifact has been embedded (for example, a synthetic watermark in one corner of every positive case), then evaluate on a test set from which the watermark is removed; a map that genuinely captures causal structure should ignore the watermark and keep accuracy, whereas an association map will depend on it and show a sharp drop.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that 'causality signals' need not be learned from interventions or expert graphs but can be read off from the statistical structure of a trained network: features that co-occur consistently in medical images carry weak causal information, and packaging that co-occurrence into a 'causality map' lets a CNN enhance the feature maps that matter, suppress those that do not, and thereby classify better while focusing on the relevant image regions. The thesis presents this as the thread linking its main results: activation maximization fails on medical images because it cannot separate class-relevant structure from anatomical regularity; prototypical part learning succeeds because it grounds decisions in radiologically meaningful parts; the causality-driven modules improve accuracy, explanation quality, and one-shot learning; CROCODILE shows that injecting the causality map as a task prior, alongside contrastive disentanglement of disease and domain branches, yields robust generalization across datasets and to out-of-distribution data; and CoCoReco shows that the same co-occurrence mechanism, dressed as biological context, improves recognition without added trainable parameters.

Load-bearing premise

The whole causality story rests on identifying feature co-occurrence with causal signal: the thesis never shows that the maps track the image-generating process rather than mere correlation, so if co-occurrence is only association, the claimed robustness and explanation improvements reduce to attention-style feature enhancement.

Editorial extensions

If this is right

  • A CNN equipped with the causality-factors module should classify more accurately and produce saliency that highlights disease-relevant image regions, improving clinician trust without retraining the backbone.
  • CROCODILE-style dual-branch training with a causality-map prior should transfer across medical domains and hold performance on out-of-distribution data better than standard fine-tuned CNNs.
  • Prototypical part learning is a viable ante-hoc explanation for breast mass classification, with prototypes that a radiologist judges clinically aligned, unlike activation maximization, which loses interpretability as accuracy rises.
  • The context-aware attention block of CoCoReco can be dropped into standard feed-forward architectures to improve recognition with no additional trainable parameters.
  • The thesis's XAI-causality scaffold gives future work a shared vocabulary, organizing the field into critics of XAI under the causality lens, XAI for causality, and causality for XAI.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If co-occurrence is sufficient as a causal proxy, the thesis's recipe removes the need for causal discovery or intervention data in a whole class of medical imaging tasks; a natural test is to pit the causality-map module against a method using an explicit causal graph on the same out-of-distribution benchmark.
  • Because the causality map is a statistic of the training distribution, it will absorb any spurious correlation that is stable in training data, so the module's robustness may fail precisely on the shortcut it learned; an intervention experiment that removes a planted co-occurring artifact from the test set would separate causal from associational behavior.
  • The same co-occurrence machinery could plausibly transfer beyond radiology to histopathology, fundus photography, and dermoscopy, wherever lesions and tissue context co-occur, and it could be paired with counterfactual generation to give per-case explanations that distinguish causes from correlates.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. This thesis, a cumulative PhD dissertation posted on arXiv, argues for 'human-aligned deep learning' along three axes: explainability, causality, and biological inspiration. Two empirical chapters evaluate XAI methods for medical classifiers: activation maximization (AM), which the author finds to produce uninterpretable visualizations for chest X-ray models and motivates a set of novel image-prior losses; and ProtoPNet for breast mass classification, validated with patient-stratified cross-validation, a genuinely held-out test set, and expert radiologist feedback. A systematic review chapter proposes a scaffold that organizes the XAI-causality literature. The core methodological contribution is a family of 'causality-driven' modules (Chapters 6-8) built on the notion of a 'causality map,' operationalized as feature co-occurrence statistics in CNN feature maps, and deployed in the CROCODILE robustness framework and the CoCoReco biologically inspired network. The load-bearing claim, stated in Section 1.2.2, is that these modules automatically discover and exploit 'weak causal signals' without a priori knowledge, improving classification, interpretability, and out-of-distribution robustness.

Significance. The thesis is honestly reported and contains genuinely solid components. Chapter 4 stands out: its evaluation protocol (patient-stratified 5-fold CV plus a held-out test set, explicitly fixing the original ProtoPNet evaluation flaw) and its single-radiologist clinical assessment provide a valuable template for explainable-by-design medical imaging research, and code is shipped for this chapter. Chapter 3's negative results and Chapter 5's systematic scaffold are useful contributions. The significance of the central causality claim, however, is conditional on an identification step—linking feature co-occurrence to causation—that the manuscript never justifies. If the causal framing fails, the modules reduce to associative attention-like enhancements and the mechanistic robustness story is weakened, even if the reported accuracy numbers survive. As an engineering contribution the work is real but modest; as a demonstration that weak causal signals can be exploited without prior knowledge, the claim is, as written, unsupported.

major comments (2)
  1. [§6.1, §1.2.2, §7.1.5, §8.1.2] The load-bearing claim of the thesis, stated in Section 1.2.2, is that the proposed modules 'automatically discover and exploit weak causal signals in medical images without requiring prior knowledge' and that this 'improves classification and produces more robust predictions by focusing on the relevant parts of the image.' The mechanism behind this claim is the 'causality map,' defined in Section 6.1 as feature co-occurrence statistics extracted from CNN feature maps, on which the causality-factors extraction module (§6.2.1), CROCODILE's prior injection (§7.1.5), and CoCoReco's Contextual Attention Block (§8.1.2) all build. The manuscript never provides the identification step that would connect co-occurrence to causation: there is no structural causal model, no do-calculus derivation, no intervention on the image-generating process, and no test distinguishing a genuinely causal feature from a merely associated one. This gap is consequential for the causal framing because the thesis itself, in Sections 2.4.2 and 2.4.3, correctly establishes that association is not causation and that shortcut learning arises from spurious correlations: a scanner artifact, a demographic attribute, or a hospital-specific marker can co-occur with a disease label without causing it, and a module that upweights such co-occurring features would amplify the shortcut rather than remove it. The robustness story of Chapter 7 and the abstract's domain-generalization claim inherit this premise. Two repairs are feasible within the manuscript's scope: (a) add an identification test, such as showing that the discovered maps track clinically known causal features, or that removing spurious co-occurrences leaves performance intact, or a capacity-matched comparison against a purely associative attention baseline; or (b) reframe the contribution as attention-like co-occurrence feature enhancement and remove the causal terminology accordingly. Without one of these, the central mechanistic claim is unsupported even if the reported accuracies hold.
  2. [§3.3.4] Key finding (i), summarized as 'accuracy rises, interpretability falls,' is drawn from Section 3.3.4, where two training regimes (transfer learning vs. end-to-end fine-tuning) are compared on two binary tasks (finding vs. no-finding and male vs. female) from ChestX-ray14. The accuracies are reported as single values with no error bars or multiple seeds (e.g., 66.76% vs. 68.98%; 79.79% vs. 95.81%), and the interpretability judgment is made by the author's visual inspection of the AM images in Figure 3.6 rather than by any quantitative interpretability measure, user study, or external rater. Since the chapter concludes a general trade-off from this comparison, the evidence base is thin. Reporting variance across seeds, matching the training budgets of the two regimes more carefully, and adding a quantifiable proxy for AM interpretability (for example, a distance-based measure of deviation from a prototypical medical image, or ratings from multiple observers) would strengthen the claim; alternatively, the conclusion should be explicitly scoped to the two configurations studied.
minor comments (6)
  1. [§2.1.2] 'ViTss divide an image' is a typo for 'ViTs divide an image.'
  2. [Eq. (3.6)] The weights λ1–λ4 of the proposed medical-image AM priors are not reported, and the text only states that tuning 'proved to be hard.' For a chapter whose principal result is negative, reporting the attempted ranges and final values would materially aid reproducibility.
  3. [Eq. (3.6) and following list] The loss components are introduced in the order Classification, Symmetry, Histogram, Noise, Frequency, but appear in Eq. (3.6) as LC, LH, LN, LS, LF; aligning the order would prevent confusion.
  4. [§6.2.1] The causality-factors extraction module introduces free parameters such as co-occurrence thresholds and the number of factors; their values and an ablation over them should be reported so that the results are reproducible and the module's sensitivity is visible.
  5. [§7.3] The results and discussion supporting the abstract's domain-generalization claim are summarized very briefly in the thesis; reproducing the key quantitative tables would allow the reader to verify the in-distribution and OOD claims directly.
  6. [§9.2.1] The heading 'Exod-IA: the Forbidden One' is informal for a formal dissertation; a neutral heading (e.g., 'Exod-IA: toward a clinical decision support system') would be more appropriate.

Circularity Check

1 steps flagged · score 6.0 of 10

The central 'causality' claim reduces to a definition: feature co-occurrence is renamed as a 'causality map', so the paper's key finding that weak causal signals can be exploited is true by construction rather than by derivation.

  1. self definitional [Section 1.2.2 (Part Two: Causality, Contributions in Chapter 6); echoed in abstract key finding (iv)]
    "Our main contribution in this chapter is to propose a novel method to automatically discover and exploit weak causal signals in medical images without requiring prior knowledge and use them to enhance CNN classifiers. To enable 'causality-driven' CNNs, we operationalize the computation of feature co-occurrence into the concept of causality map, design a new causality-factors extraction module, and present a novel, attention-inspired scheme based on feature map enhancement."

    The claimed output (discovery and exploitation of 'weak causal signals') is exactly the input (feature co-occurrence statistics) under a new name. The 'causality map' is not derived from a causal model; it is defined in the same sentence as the operationalization of co-occurrence. Therefore the statement that the module exploits causal signals is not an empirical finding: the predicate 'causal' is attached to the co-occurrence statistic by definition. No SCM, intervention, do-calculus derivation, or causal identification step is supplied to move from association to causation, so any observed accuracy or robustness gain supports only the co-occurrence-based attention mechanism, not the causal interpretation.

full rationale

The thesis has substantial independent empirical content: Chapter 4 benchmarks ProtoPNet on CBIS-DDSM with held-out test data and clinical feedback, Chapters 6-7 report accuracy, ablation, and OOD experiments on public medical datasets, and code is referenced. The author's self-citations are to peer-reviewed publications and are not used to replace evidence; they do not by themselves create circularity. The single genuine circularity is the co-occurrence-to-causality identification. Because 'causality map' is defined as the operationalization of feature co-occurrence, the headline claim that 'weak causal signals can be leveraged without a priori information' is equivalent to saying that co-occurrence statistics can be computed and used—this is true by definition. The prediction of improved classification is empirical and survives separately, but the mechanistic claim that the modules are 'causality-driven' reduces to a renaming of association. This is partial circularity in the central claim, not a fully circular derivation of the accuracy results, hence a score of 6 rather than 8 or 10.

Assumptions & free parameters 3 free parameters · 5 assumptions · 5 invented entities

The thesis is empirical, so the ledger is light on fitted constants and heavy on stated domain assumptions. The genuine free parameters are the AM prior weights and the causality-map hyperparameters; the ProtoPNet coefficients are standard dataset-specific tuning. The invented entities are software modules rather than physical postulates, and none carries an independent falsifiable handle beyond downstream task accuracy. The causal semantics of co-occurrence is the single most consequential assumption in the ledger.

free parameters (3)
  • AM prior loss weights lambda1-lambda4 (Eq. 3.6) = Final values not reported; hand-tuned on the target dataset and abandoned after failing to yield interpretable outputs
    Chapter 3.3.5 combines classification, histogram, symmetry, noise, and frequency losses; the author states that finding a working equilibrium 'proved to be hard' and no final weight set is given, so the reported negative result is conditional on these unstated constants.
  • Causality map hyperparameters (co-occurrence thresholds, number of factors) = Not visible in the available text (Chapter 6 experimental section truncated)
    The causality-factors extraction module computes co-occurrence statistics from CNN features; its thresholds and factor counts are selected to optimize validation accuracy and are not available for scrutiny, so the headline gains of Chapter 6 are conditional on them.
  • ProtoPNet loss coefficients clst, sep and prototype count = clst in {0.6, 0.8, 0.9}, sep in {-0.1, -0.08, -0.05}, prototypes per class in {5, 20, 40}
    Table 4.1 defines the random search domain; the clustering and separation coefficients directly shape which prototypes are learned, and the chosen values are dataset-specific fits via five-fold CV.
assumptions (5)
  • domain assumption Co-occurrence of features in CNN feature maps constitutes a weak causal signal exploitable without a priori knowledge.
    Chapter 6.1 'Causality Signals in Images' defines the causality map operationally as feature co-occurrence; Chapters 7 and 8 build on it. No causal identification (intervention, counterfactual, or back-door adjustment over actual image variables) is provided, so the causal reading of association data is assumed.
  • domain assumption ImageNet-pretrained backbone features transfer to grayscale medical images.
    VGG16 and ResNet18 backbones pretrained on ImageNet are fine-tuned on chest X-rays (Chapter 3.3.1), mammograms (Chapter 4.3.3), and used in Chapters 6 and 8; the baselines and gains all sit on this transfer assumption.
  • domain assumption CXR structural priors (bilateral symmetry, GMM intensity histogram, Poisson quantum mottle, FFT magnitude spectrum) are valid generative constraints for activation maximization.
    Chapter 3.3.5, Eqs. 3.6-3.10; the author reports these priors did not produce interpretable AM outputs, so this assumption is load-bearing for the chapter's mitigation strategy and was empirically unsupported in its outcome.
  • standard math Pearl's SCM and back-door adjustment formalism is the correct language for the disease/domain disentanglement.
    Chapter 2.4 introduces SCMs, do-calculus, and back-door adjustment; Chapter 7.1.1 frames CROCODILE in these terms. This is standard accepted background, invoked rather than derived.
  • domain assumption Grad-CAM heatmaps are a trustworthy proxy for what the model sees in the qualitative evaluations.
    Grad-CAM is used for qualitative XAI evaluation in Chapters 6 and 8, even though Section 2.3.3 documents Grad-CAM's known localization failures; relying on a method the thesis itself criticizes is a fragile evaluation premise.
invented entities (5)
  • Causality map (feature co-occurrence map)
    purpose: Encodes how often CNN features co-occur and guides feature-map enhancement toward regions claimed to carry weak causal signal.
    Introduced in Chapter 6.1; validated only through downstream accuracy and explanation gains on the same datasets, with no external falsifiable handle separating co-occurrence from causation.
  • Causality-factors extraction module
    purpose: Extracts the factors that drive the causality map and applies an attention-inspired enhancement scheme to CNN feature maps.
    A paper-internal component assessed only within Chapter 6's own experiments; no independent benchmark or external constraint.
  • Relational scorer module (CROCODILE)
    purpose: Compares latent representations from the disease branch and the domain branch to penalize spurious cross-domain associations.
    Chapter 7.1; the 'causal' vs 'spurious' split is defined by the authors' own branch assignment, and OOD accuracy is the only check, so the module has no independent falsifiable handle.
  • Contextual Attention Block (CAB)
    purpose: Models feature co-occurrence (context) at network bottlenecks without adding trainable parameters, inspired by biological visual context.
    Chapter 8.1.2; evaluated only inside CoCoReco on the author's chosen tasks, with no external context-modeling benchmark.
  • Exod-IA (future system)
    purpose: Sketched in Future Directions as a planned clinical deployment of explainable, causal AI.
    Section 9.2.1 explicitly labels this as future work; it contributes no current scientific claim, so it is listed for completeness only.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Human-aligned Deep Learning: Explainability, Causality, and Biological Inspiration." pith.science (2026). https://pith.science/paper/NJ7BJ3LP

@misc{pith2026250413717,
  author       = {Pith},
  title        = {Pith review of: Human-aligned Deep Learning: Explainability, Causality, and Biological Inspiration},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NJ7BJ3LP}},
  note         = {Machine review of arXiv:2504.13717}
}
read the original abstract

This work aligns deep learning (DL) with human reasoning capabilities and needs to enable more efficient, interpretable, and robust image classification. We approach this from three perspectives: explainability, causality, and biological vision. Introduction and background open this work before diving into operative chapters. First, we assess neural networks' visualization techniques for medical images and validate an explainable-by-design method for breast mass classification. A comprehensive review at the intersection of XAI and causality follows, where we introduce a general scaffold to organize past and future research, laying the groundwork for our second perspective. In the causality direction, we propose novel modules that exploit feature co-occurrence in medical images, leading to more effective and explainable predictions. We further introduce CROCODILE, a general framework that integrates causal concepts, contrastive learning, feature disentanglement, and prior knowledge to enhance generalization. Lastly, we explore biological vision, examining how humans recognize objects, and propose CoCoReco, a connectivity-inspired network with context-aware attention mechanisms. Overall, our key findings include: (i) simple activation maximization lacks insight for medical imaging DL models; (ii) prototypical-part learning is effective and radiologically aligned; (iii) XAI and causal ML are deeply connected; (iv) weak causal signals can be leveraged without a priori information to improve performance and interpretability; (v) our framework generalizes across medical domains and out-of-distribution data; (vi) incorporating biological circuit motifs improves human-aligned recognition. This work contributes toward human-aligned DL and highlights pathways to bridge the gap between research and clinical adoption, with implications for improved trust, diagnostic accuracy, and safe deployment.

Figures

Figures reproduced from arXiv: 2504.13717 by the authors.

Figure 1.1
Figure 1.1. Dissertation Road-map. 2 [PITH_FULL_IMAGE:figures/full_fig_p014_1_1.png] view at source ↗
Figure 1.2
Figure 1.2. Chapters’ Organization. Click these numbers to skip to Chapter: 2, 3, 4, 5, 6, 7, 8, 9. In the following, we highlight this research’s original contributions to those fields, including new findings, methodologies, and theoretical advancements. 1.2.1 Part One: Explainability In the first part of the thesis (Chapters 3 and 4), we focus on the explainability of DL models, investigating the need for interpretations aris… view at source ↗
Figure 2.1
Figure 2.1. Post-hoc and Ante-hoc XAI. In post-hoc methods, the outcome of a black-box model is passed through an additional, surrogate, interpretable model that generates an explanation. Ante-hoc methods avoid using a black-box model in the first place and instead build an inherently transparent model. So, together with an outcome, it already provides an explanation - the reasoning process is transparent. in the same way the m… view at source ↗
Figures from the paper (47 more)
Figure 2.2
Figure 2.2. Figure 2.2: Class Activation Maps for five popular categories of ImageNet. For each class, three examples are shown. The Grad-CAM maps are superimposed on the original image as heatmaps: higher activation values are represented in dark red; the lower the value, the bluish the pi…
Figure 2
Figure 2. Figure 2: ). ProtoPNet dissects images into parts, learns a set of prototypes for each [PITH_FULL_IMAGE:figures/full_fig_p036_2.png]
Figure 2.3
Figure 2.3. Figure 2.3: Difference between Prototypical Learning and Prototypical Part Learning. While the former (left panel) compares one whole image to another whole image, the latter (right panel) seeks the relevant parts of the input image that led to a specific class prediction. In ot…
Figure 2.4
Figure 2.4. Figure 2.4: Feature Visualization. The idea behind Activation Maximization (AM) is to leverage the differentiable nature of neural networks w.r.t. their inputs. The goal is to compute the input image that maximally activates a specific neuron (e.g., the ’knot’ class) in a traine…
Figure 2.5
Figure 2.5. Figure 2.5: Examples of directed graphs: (a) directed cyclic graph, (b) directed acyclic graph (DAG). 2.4.3 How? Causal Modelling: From Association to Causation Directed Acyclic Graph (DAG) Much of the discussion of causality and qualitative modeling is occupied by DAGs [316]. D…
Figure 2.6
Figure 2.6. Figure 2.6: Example of a fully specified BN which models the probability of observing wet grass. In this (simplified) real-world scenario, grass can be wet either by turning on a sprinkler or by rainfall, and both can be influenced by the presence of clouds in the sky. factuals …
Figure 2.7
Figure 2.7. Figure 2.7: The ventral and dorsal visual pathways in human vision. Adapted from https://www. brainfacts.org/ of the Society for Neuroscience (2017). medical image segmentation. [352] propose using probabilistic causal models to gen￾erate high-fidelity image counterfactuals. All…
Figure 3.1
Figure 3.1. Figure 3.1: Feature visualization through Activation Maximization on selected output neurons of the Inception-V3 model pre-trained on ImageNet (1000 classes). Each panel shows the obtained image that maximally activates the output neuron corresponding to that specific class (out…
Figure 3.2
Figure 3.2. Figure 3.2: Exaggerating image size and number of iterations. These visualizations are obtained from Inception-V3 with 768 × 768 input image and 10000 AM iterations. The large image size of the input (×3.5 w.r.t original) results in a compositional feature visualization, where s…
Figure 3.3
Figure 3.3. Figure 3.3: AM exposes dataset biases and shortcuts (part 1). These visualizations, obtained from an Inception-V3 model with 1000 AM iterations, are exemplars of the spurious features that can be associated with the object category due to dataset bias. The AM optimization for th…
Figure 3.4
Figure 3.4. Figure 3.4: AM exposes dataset biases and shortcuts (part 2). This visualization is another example of spurious features associated with the object category due to bias. The AM optimization for the ’llama’ class produced feature visualizations showing structures of the diamond-s…
Figure 3.5
Figure 3.5. Figure 3.5: Feature visualizations obtained by repeating the optimization process six times (columns) for each image class (rows) with the same learning settings. 50 [PITH_FULL_IMAGE:figures/full_fig_p062_3_5.png]
Figure 3.6
Figure 3.6. Figure 3.6: Feature visualizations of output neurons on VGG16 models for radiology finding and patient gender classification from chest X-ray images. The top half of the figure refers to a TL regime, where only the last fully connected layer is trained and the rest of the networ…
Figure 3.7
Figure 3.7. Figure 3.7: The average chest X-ray (CXR) image and some CXRs from the training set of the NIH’s ChestXRay14 dataset. The average image is obtained by computing the mean pixel intensities over more than 86000 image samples, and this visualization helps understand how practically…
Figure 3.8
Figure 3.8. Figure 3.8: Exploring the space of AM hyperparameters is challenging for medical images. The output greatly varies with the chosen values of LR, WD, and jittering and is hardly interpretable to humans. The example shown represents the AM images of the female neuron of the VGG16 …
Figure 4.1
Figure 4.1. Figure 4.1: Examples of images from the original CBIS-DDSM dataset that were removed due to arti￾facts. (a)-(b): annotation spot next to or within the mass; (c): white-band artifact; (d) horizontal￾pattern artifact. or within the mass region ( [PITH_FULL_IMAGE:figures/full_fig_…
Figure 4.2
Figure 4.2. Figure 4.2: Learning the Prototypes. Where do those prototypes come from? The training process of ProtoPNet consists of learning a meaningful latent representation where characteristic patches of training images become centroids in a clustering process, that is, prototypes. marg…
Figure 4.3
Figure 4.3. Figure 4.3: Inference process through ProtoPNet: classification of a breast mass by means of the acti￾vation of pre-learned prototypes within the test image. 4.3.3 Experimental Settings As for the CNN block of ProtoPNet, the residual network ResNet18 with weights pre￾trained on …
Figure 4.4
Figure 4.4. Figure 4.4: Five-fold Cross Validation. The train set was split in 5 parts, one was used as the internal validation and the remaining 4 as the internal training. All the splits were patient based, in other words, masses of the same patients were in the same “internal” set (train…
Figure 4.5
Figure 4.5. Figure 4.5: Average accuracy curves across the five CV folds for the selected ProtoPNet model’s con￾figuration. Shaded regions represent ±1 · std interval for each epoch [PITH_FULL_IMAGE:figures/full_fig_p080_4_5.png]
Figure 4.6
Figure 4.6. Figure 4.6: The test image of a malignant mass is correctly classified as malignant by the model. Each row of this image represents the activation process of a certain prototype. In the first column there is the patch found on the test image, in the second column the activated p…
Figure 4.7
Figure 4.7. Figure 4.7: An experienced radiologist provided clinical feedback on the goodness of the produced explanations. radiologist with 16 years of experience for a clinical viewpoint on the outputs of the selected model on a random subset of test images (15 benign, 15 malignant) to pr…
Figure 5.1
Figure 5.1. Figure 5.1: Graphical Abstract: Reviewing the literature to uncover how eXplainable Artificial Intelli￾gence (XAI) and causality are related - the three main perspectives. 5.1 Ancient roots The study of causation and explanation can be traced back to the ancient Greek philoso￾ph…
Figure 5
Figure 5. Figure 5: shows the process of identification, screening, eligibility, and inclusion [PITH_FULL_IMAGE:figures/full_fig_p092_5.png]
Figure 5.2
Figure 5.2. Figure 5.2: Flowchart of the study collection process, from identification, through screening, to eligibil￾ity and inclusion. 5.3.3 Keywords’ co-occurrence analysis Regarding the high-level analysis of the final cohort of records, we constructed a bib￾liometric network of articl…
Figure 5.3
Figure 5.3. Figure 5.3: Bibliometric network of papers’ keywords for the cohort of publications included in the review. a b [PITH_FULL_IMAGE:figures/full_fig_p096_5_3.png]
Figure 5.4
Figure 5.4. Figure 5.4: The isolated connections from [PITH_FULL_IMAGE:figures/full_fig_p096_5_4.png]
Figure 5.5
Figure 5.5. Figure 5.5: The included studies are classified according to the three main perspectives on how causality and XAI may be related: Critics to XAI under the causality lens, XAI for causality, and Causality for XAI. Next to each of them, are the possible sub-clusters. 5.5 Results t…
Figure 6.1
Figure 6.1. Figure 6.1: Zoomed-in visualizations of a sample causality map computed with Eq. 6.1 on 512 feature maps extracted from an input image. (a) 512 × 512 original causality map; (b) 256 × 256 zoom￾in of (a); (c) 64 × 64 zoom-in of (b), where dashed circles indicate exemplar elements…
Figure 6.2
Figure 6.2. Figure 6.2: Overview of the different settings investigated in this work: Baseline, Cat, and the proposed Mulcat. Assuming k = 6 feature maps as an example, the tensor F of feature maps that are obtained from a CNN just before the classifier can be either flattened and used as t…
Figure 6.3
Figure 6.3. Figure 6.3: The internals of the proposed causality factors extractor block of [PITH_FULL_IMAGE:figures/full_fig_p119_6_3.png]
Figure 6.4
Figure 6.4. Figure 6.4: Some benign and malignant samples from the utilized BreakHis dataset. This image is best seen in color. • bool. The factors undergo boolean thresholding where all the non-zero factors are assigned a new weight of 1 and 0 otherwise. As a result, this choice is more co…
Figure 6.5
Figure 6.5. Figure 6.5: Some no-tumour and cancerous samples from the utilized PI-CAI dataset. PI-CAI Dataset From the available 1500 acquisitions, we only selected T2-weighted (T2w) images. Within this cohort of patients and respective scans, some cases didn’t have any tumors (i.e., they h…
Figure 6.6
Figure 6.6. Figure 6.6: Visual assessment of class activation maps for different malignant cases of the BreakHis dataset. Each row represents a different biopsy slide, and the columns represent the Grad-CAM outputs for the baseline model, the Cat option, and all the proposed Mulcat variants…
Figure 6.7
Figure 6.7. Figure 6.7: Visual assessment of class activation maps for cancerous cases of the PI-CAI dataset. Each row represents a different scan, and the columns represent the Grad-CAM outputs for the baseline model, the Cat option, and all the proposed Mulcat variants. ResNet18 is assume…
Figure 6.8
Figure 6.8. Figure 6.8: Visual assessment of class activation maps for no-tumor cases of the PI-CAI dataset. Each row represents a different scan, and the columns represent the Grad-CAM outputs for the baseline model, the Cat option, and all the proposed Mulcat variants. ResNet18 is assumed…
Figure 7.1
Figure 7.1. Figure 7.1: A causal view on classifying medical images I coming from different domains D for the presence of diseases Y. By applying the latent causal intervention (hammer), the backdoor path through the spurious features is cut off. 7.1 Methodology 7.1.1 A Causal Viewpoint on …
Figure 7.2
Figure 7.2. Figure 7.2: CROCODILE involves two branches to learn robust, invariant features for predicting the labels from medical images (e.g., multi-label findings from CXRs) while disregarding confounding features. We disentangle causal features determining the label from spurious featur…
Figure 7.3
Figure 7.3. Figure 7.3: Our Relational Scorer stratifies and concatenates every combination of causal and spurious features across both tasks. With a fully connected layer and a consecutive sigmoid(·), it maps each pair to a relational score between 0 and 1. We use an MSE loss regressing th…
Figure 7.4
Figure 7.4. Figure 7.4: Causal graphical model among the CXR findings of interest (blue) and the ground-truth causality map defined from that graph. Gray boxes represent additional findings or risk factors (not investigated in this study) that might be associated with the desired ones. 7.2 …
Figure 8.1
Figure 8.1. Figure 8.1: Overview of our Connectivity-inspired Context aware Recognition network. The internals and rationale of the CAB module are presented in Sec. 8.1.2 and [PITH_FULL_IMAGE:figures/full_fig_p151_8_1.png]
Figure 8.2
Figure 8.2. Figure 8.2: Our Contextual Attention Block (CAB) integrated into a general feed-forward network. As shown, CAB is placed at the convolutional bottleneck of the model. Given intermediate feature maps, the module computes corresponding contextual attention scores through a rectifi…
Figure 8.3
Figure 8.3. Figure 8.3: GradCAM activations for three test images for CoCoReco, ablation, and baseline models. The left panel (blue) shows the outputs when the last convolutional layer before the classifier is chosen as the target layer for the GradCAM computation. The panel on the right (o…
Figure 8
Figure 8. Figure 8: (left panel) shows the models’ outputs on three example test images when [PITH_FULL_IMAGE:figures/full_fig_p157_8.png]
Figure 9.1
Figure 9.1. Figure 9.1: Exod-IA: the Forbidden One. We could then design a hybrid loss function that combines the standard CE loss used in classification with the prototype similarity loss. This way, the Transformer layers would be trained to not only classify the input based on embeddings …
Figure 9.2
Figure 9.2. Figure 9.2: Extended causal graphical model of the data-generating process of CXR findings from patient-specific attributes and demographic information. X-ray image higher probability of KVP EXPOSURE TIME higher TUBE CURRENT copper:higher molybdenum:higher FILTER GRID/COLLIMATOR…
Figure 9.3
Figure 9.3. Figure 9.3 [PITH_FULL_IMAGE:figures/full_fig_p168_9_3.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Causal Transfer in Medical Image Analysis

    cs.CV 2026-03 accept novelty 5.0 of 10

    Causal Transfer Learning unifies structural causal models, invariant risk minimisation and counterfactuals with transfer learning to produce domain-robust medical image models.

Reference graph

Works this paper leans on

39 extracted references · 29 canonical work pages · cited by 1 Pith paper

  1. [482]

    Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains

    Mingjun Xu, Lingyun Qin, Weijie Chen, Shiliang Pu, and Lei Zhang. Multi-view adversarial discriminator: Mine the non-causal factors for object detection in unseen domains. InProceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 8103–8112, 2023

  2. [483]

    Novel robust automatic brain-tumor detection and segmentation using magnetic resonance imaging.IEEE Sensors Journal, 2024

    Mingyang Xu, Limei Guo, and Hsiao-Chun Wu. Novel robust automatic brain-tumor detection and segmentation using magnetic resonance imaging.IEEE Sensors Journal, 2024

  3. [484]

    Cine cardiac mri reconstruction using a convolutional recurrent network with refinement

    Yuyang Xue, Yuning Du, Gianluca Carloni, Eva Pachetti, Connor Jordan, and Sotirios A Tsaftaris. Cine cardiac mri reconstruction using a convolutional recurrent network with refinement. In International Workshop on Statistical Atlases and Computational Models of the Heart, pages 421–432. Springer, 2023

  4. [485]

    Epvt: Environment-aware prompt vision transformer for domain generalization in skin lesion recognition

    Siyuan Yan, Chi Liu, Zhen Yu, Lie Ju, Dwarikanath Mahapatra, Victoria Mar, Monika Janda, Peter Soyer, and Zongyuan Ge. Epvt: Environment-aware prompt vision transformer for domain generalization in skin lesion recognition. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 249–259. Springer, 2023

  5. [486]

    Melanoma recognition via visual attention

    Yiqi Yan, Jeremy Kawahara, and Ghassan Hamarneh. Melanoma recognition via visual attention. InInformation Processing in Medical Imaging: 26th International Conference, IPMI 2019, Hong Kong, China, June 2–7, 2019, Proceedings 26, pages 793–804. Springer, 2019

  6. [487]

    Lung nodule detection algorithm based on rank correlation causal structure learning.Expert Systems with Applications, 216:119381, 2023

    Jing Yang, Liufeng Jiang, Kai Xie, Qiqi Chen, and Aiguo Wang. Lung nodule detection algorithm based on rank correlation causal structure learning.Expert Systems with Applications, 216:119381, 2023

  7. [488]

    Algorithmic foundation of deep x-risk optimization.arXiv preprint arXiv:2206.00439, 2022

    Tianbao Yang. Algorithmic foundation of deep x-risk optimization.arXiv preprint arXiv:2206.00439, 2022

  8. [489]

    An interpretable unsupervised bayesian network model for fault detection and diagnosis.Control Engineering Practice, 127:105304, 2022

    Wei-Ting Yang, Marco S Reis, Valeria Borodin, Michel Juge, and Agnès Roussy. An interpretable unsupervised bayesian network model for fault detection and diagnosis.Control Engineering Practice, 127:105304, 2022

Show all 39 references
  1. [490]

    Yi Yang, Yueting Zhuang, and Yunhe Pan. Multiple knowledge representation for big data artificial intelligence: framework, applications, and case studies.Frontiers of Information Technology & Electronic Engineering, 22(12):1551–1558, 2021

  2. [491]

    Unifying causal representation learning with the invariance principle.arXiv preprint arXiv:2409.02772, 2024

    Dingling Yao, Dario Rancati, Riccardo Cadei, Marco Fumero, and Francesco Locatello. Unifying causal representation learning with the invariance principle.arXiv preprint arXiv:2409.02772, 2024

  3. [492]

    Causal markov blanket representation learning for out-of-distribution generalization

    Naiyu Yin, Hanjing Wang, Tian Gao, Amit Dhurandhar, and Qiang Ji. Causal markov blanket representation learning for out-of-distribution generalization. InCausal Representation Learning Workshop at NeurIPS 2023, 2023

  4. [493]

    Understanding neural networks through deep visualization.arXiv preprint arXiv:1506.06579, 2015

    Jason Yosinski, Jeff Clune, Anh Nguyen, Thomas Fuchs, and Hod Lipson. Understanding neural networks through deep visualization.arXiv preprint arXiv:1506.06579, 2015

  5. [494]

    Elsevier Health Sciences, 2013

    Barbara Young, Geraldine O’Dowd, and Phillip Woodford.Wheater’s functional histology E-Book: a text and colour atlas. Elsevier Health Sciences, 2013

  6. [495]

    A super-resolution network for medical imaging via transformation analysis of wavelet multi-resolution.Neural Networks, 166:162–173, 2023

    Yue Yu, Kun She, Jinhua Liu, Xiao Cai, Kaibo Shi, and Oh-Min Kwon. A super-resolution network for medical imaging via transformation analysis of wavelet multi-resolution.Neural Networks, 166:162–173, 2023

  7. [496]

    Po-1794 features robustness in the radiomic workflow: the impact of software choice on feature variability

    Mattia Zaffaroni, Stefania Carloni, Gianluca V olpe, Cristina Garibaldi, Giulia Marvaso, et al. Po-1794 features robustness in the radiomic workflow: the impact of software choice on feature variability. Radiotherapy and Oncology, 161:S1519–S1520, 2021

  8. [497]

    An interpretable approach to automated severity scoring in pelvic trauma

    Anna Zapaishchykova, David Dreizin, Zhaoshuo Li, Jie Ying Wu, Shahrooz Faghihroohi, and Mathias Unberath. An interpretable approach to automated severity scoring in pelvic trauma. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conf...

  9. [498]

    Scientific exploration and explainable artificial intelligence.Minds and Machines, 32(1):219–239, 2022

    Carlos Zednik and Hannes Boelsen. Scientific exploration and explainable artificial intelligence.Minds and Machines, 32(1):219–239, 2022

  10. [499]

    A survey on causal reinforcement learning.arXiv preprint arXiv:2302.05209, 2023

    Yan Zeng, Ruichu Cai, Fuchun Sun, Libo Huang, and Zhifeng Hao. A survey on causal reinforcement learning.arXiv preprint arXiv:2302.05209, 2023

  11. [500]

    Review of shape representation and description techniques.Pattern Recognition, 37(1):1–19, 2004

    Dengsheng Zhang and Guojun Lu. Review of shape representation and description techniques.Pattern Recognition, 37(1):1–19, 2004

  12. [501]

    Causal intervention for weakly-supervised semantic segmentation.Advances in Neural Information Processing Systems, 33:655–666, 2020

    Dong Zhang, Hanwang Zhang, Jinhui Tang, Xian-Sheng Hua, and Qianru Sun. Causal intervention for weakly-supervised semantic segmentation.Advances in Neural Information Processing Systems, 33:655–666, 2020

  13. [502]

    Identifiability guarantees for causal disentanglement from soft interventions

    Jiaqi Zhang, Kristjan Greenewald, Chandler Squires, Akash Srivastava, Karthikeyan Shanmugam, and Caroline Uhler. Identifiability guarantees for causal disentanglement from soft interventions. Advances in Neural Information Processing Systems, 36, 2024

  14. [503]

    Towards causal foundation model: on duality between causal inference and attention.arXiv preprint arXiv:2310.00809, 2023

    Jiaqi Zhang, Joel Jennings, Cheng Zhang, and Chao Ma. Towards causal foundation model: on duality between causal inference and attention.arXiv preprint arXiv:2310.00809, 2023

  15. [504]

    Learning towards synchronous network memorizability and generalizability for continual segmentation across multiple sites

    Jingyang Zhang, Peng Xue, Ran Gu, Yuning Gu, Mianxin Liu, Yongsheng Pan, Zhiming Cui, Jiawei Huang, Lei Ma, and Dinggang Shen. Learning towards synchronous network memorizability and generalizability for continual segmentation across multiple sites. InInternational Conference ...

  16. [505]

    Coping with label shift via distributionally robust optimisation.arXiv preprint arXiv:2010.12230, 2020

    Jingzhao Zhang, Aditya Menon, Andreas Veit, Srinadh Bhojanapalli, Sanjiv Kumar, and Suvrit Sra. Coping with label shift via distributionally robust optimisation.arXiv preprint arXiv:2010.12230, 2020

  17. [506]

    Causal representation learning from multiple distributions: A general setting.arXiv preprint arXiv:2402.05052, 2024

    Kun Zhang, Shaoan Xie, Ignavier Ng, and Yujia Zheng. Causal representation learning from multiple distributions: A general setting.arXiv preprint arXiv:2402.05052, 2024

  18. [507]

    Putting visual object recognition in context

    Mengmi Zhang, Claire Tseng, and Gabriel Kreiman. Putting visual object recognition in context. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12985–12994, 2020

  19. [508]

    Personalized latent structure learning for recommendation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10285–10299, 2023

    Shengyu Zhang, Fuli Feng, Kun Kuang, Wenqiao Zhang, Zhou Zhao, Hongxia Yang, Tat-Seng Chua, and Fei Wu. Personalized latent structure learning for recommendation.IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(8):10285–10299, 2023

  20. [509]

    Out-of-distribution generalized dynamic graph neural network with disentangled intervention and invariance promotion.arXiv preprint arXiv:2311.14255, 2023

    Zeyang Zhang, Xin Wang, Ziwei Zhang, Haoyang Li, and Wenwu Zhu. Out-of-distribution generalized dynamic graph neural network with disentangled intervention and invariance promotion.arXiv preprint arXiv:2311.14255, 2023

  21. [510]

    Explainable contrastive multiview graph representation of brain, mind, and behavior

    Chongyue Zhao, Liang Zhan, Paul M Thompson, and Heng Huang. Explainable contrastive multiview graph representation of brain, mind, and behavior. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 356–365. Springer, 2022

  22. [511]

    Disentangling site effects with cycle-consistent adversarial autoencoder for multi-site cortical data harmonization

    Fenqiang Zhao, Zhengwang Wu, Dajiang Zhu, Tianming Liu, John Gilmore, Weili Lin, Li Wang, and Gang Li. Disentangling site effects with cycle-consistent adversarial autoencoder for multi-site cortical data harmonization. InInternational Conference on Medical Image Computing and...

  23. [512]

    Causal-learn: Causal discovery in python.Journal of Machine Learning Research, 25(60):1–8, 2024

    Yujia Zheng, Biwei Huang, Wei Chen, Joseph Ramsey, Mingming Gong, Ruichu Cai, Shohei Shimizu, Peter Spirtes, and Kun Zhang. Causal-learn: Causal discovery in python.Journal of Machine Learning Research, 25(60):1–8, 2024

  24. [513]

    On the opportunity of causal deep generative models: A survey and future directions.arXiv preprint arXiv:2301.12351, 2023

    Guanglin Zhou, Lina Yao, Xiwei Xu, Chen Wang, Liming Zhu, and Kun Zhang. On the opportunity of causal deep generative models: A survey and future directions.arXiv preprint arXiv:2301.12351, 2023

  25. [514]

    Chest radiograph disentanglement for covid-19 outcome prediction

    Lei Zhou, Joseph Bae, Huidong Liu, Gagandeep Singh, Jeremy Green, Dimitris Samaras, and Prateek Prasanna. Chest radiograph disentanglement for covid-19 outcome prediction. InMedical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, ...

  26. [515]

    Learning with domain-knowledge for generalizable prediction of alzheimer’s disease from multi-site structural mri

    Yanjie Zhou, Youhao Li, Feng Zhou, Yong Liu, and Liyun Tu. Learning with domain-knowledge for generalizable prediction of alzheimer’s disease from multi-site structural mri. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 452–461...

  27. [516]

    Counterfactual fairness by combining factual and counterfactual predictions.arXiv preprint arXiv:2409.01977, 2024

    Zeyu Zhou, Tianci Liu, Ruqi Bai, Jing Gao, Murat Kocaoglu, and David I Inouye. Counterfactual fairness by combining factual and counterfactual predictions.arXiv preprint arXiv:2409.01977, 2024

  28. [517]

    How well do feature visualizations support causal understanding of cnn activations? Advances in Neural Information Processing Systems, 34:11730–11744, 2021

    Roland S Zimmermann, Judy Borowski, Robert Geirhos, Matthias Bethge, Thomas Wallis, and Wieland Brendel. How well do feature visualizations support causal understanding of cnn activations? Advances in Neural Information Processing Systems, 34:11730–11744, 2021

  29. [518]

    Edge detection techniques-an overview.Pattern Recognition and Image Analysis: Advances in Mathematical Theory and Applications, 8(4):537–559, 1998

    Djemel Ziou and Salvatore Tabbone. Edge detection techniques-an overview.Pattern Recognition and Image Analysis: Advances in Mathematical Theory and Applications, 8(4):537–559, 1998

  30. [519]

    Mohammad Zunaed, Md Aynal Haque, and Taufiq Hasan. Learning to generalize towards unseen domains via a content-aware style invariant model for disease detection from chest x-rays.IEEE Journal of Biomedical and Health Informatics, 2024

  31. [520]

    Alex Zwanenburg. Radiomics in nuclear medicine: robustness, reproducibility, standardization, and how to avoid data analysis traps and replication crisis.European journal of nuclear medicine and molecular imaging, 46(13):2638–2655, 2019. 170

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.