PaperlayerView exact revision on arXiv

Source-linked AI summary

Explaining the Black-box Smoothly- A Counterfactual Approach

Sumedha Singla, Motahhare Eslami, Brian Pollack, Stephen Wallace, Kayhan Batmanghelich

arXiv:2101.04230v3cs.CVeess.IV

TL;DR

Medical image explanations can identify important features without showing how anatomical features relate to a classifier’s decision. This paper presents a GAN-based counterfactual explainer, and radiology residents’ understanding of classifier decisions improved significantly only with this method compared with no explanation.

  • Problem

    Saliency maps identify important features but do not explain how features in anatomical regions relate to a medical image classifier’s decision.

    Source: §1 · p. 2, §5 · p. 17

  • Method

    The paper introduces a GAN-based counterfactual explainer that progressively alters query images to move a classifier’s decision from its original class toward its negation.

    Source: §1 · p. 2, §5 · p. 16

  • Results

    The counterfactual explanation was the only method to significantly improve radiology residents’ understanding of classifier decisions compared with no explanation.

    Source: §5 · p. 17

  • Takeaways & Limitations

    The authors conclude that the classifier relied on clinically relevant radiographic features, making its decisions more transparent to end users.

    Source: §5 · pp. 16–17

    The authors acknowledge an issue with GAN-generated counterfactual explanations and leave its impact on interpretation for future work.

    Source: §5 · p. 17

Abstract

from arXiv · show

We propose a BlackBox Counterfactual Explainer, designed to explain image classification models for medical applications. Classical approaches (e.g., saliency maps) that assess feature importance do not explain "how" imaging features in important anatomical regions are relevant to the classification decision. Our framework explains the decision for a target class by gradually "exaggerating" the semantic effect of the class in a query image. We adopted a Generative Adversarial Network (GAN) to generate a progressive set of perturbations to a query image, such that the classification decision changes from its original class to its negation. We used counterfactual explanations from our framework to audit a classifier trained on a chest x-ray dataset with multiple labels. We proposed clinically-relevant quantitative metrics such as cardiothoracic ratio and the score of a healthy costophrenic recess to evaluate our explanations. We conducted a human-grounded experiment with diagnostic radiology residents to compare different styles of explanations (no explanation, saliency map, cycleGAN explanation, and our counterfactual explanation) by evaluating different aspects of explanations: (1) understandability, (2) classifier's decision justification, (3) visual quality, (d) identity preservation, and (5) overall helpfulness of an explanation to the users. Our results show that our counterfactual explanation was the only explanation method that significantly improved the users' understanding of the classifier's decision compared to the no-explanation baseline. Our metrics established a benchmark for evaluating model explanation methods in medical images. Our explanations revealed that the classifier relied on clinically relevant radiographic features for its diagnostic decisions, thus making its decision-making process more transparent to the end-user.

1. Introduction

The introduction frames counterfactual image explanations as a way to show how imaging features influence classifier decisions, beyond location-only heatmaps, and presents a progressive explainer alongside prior perturbation methods that may produce anatomically implausible images. It outlines methods for preserving anatomy and evaluating explanations in chest X-rays, including clinical metrics and a human-grounded study with radiology residents.

1. Introduction

  • Unlike location-only heatmaps, a thorough explanation should identify the imaging features in important regions and how changing them modifies the decision.The introduction argues that different diagnoses can affect the same anatomical regions, making location-only explanations inconclusive.Source: §1 · p. 2
  • The proposed counterfactual explanation shows which image features change as an input is perturbed to flip the classifier’s decision.The introduction describes a positive pleural-effusion prediction changing to negative, with changes concentrated in the lower-lobe region.Source: §1 · p. 2
  • The method generates a series of perturbations that traverse the decision boundary and uses semantic segmentation and object detection in its loss to preserve anatomy and foreign objects.The introduction also describes extending a conditional GAN to preserve small or uncommon details during image generation.Source: §1 · p. 2
  • The authors evaluated explanations with quantitative and clinical measures and conducted a comparison study with 12 radiology residents.The introduction says residents compared the proposed method with baseline models, but this passage does not report study outcomes.Source: §1 · pp. 2–3

1.1. Related work

  • Prior perturbation methods can introduce anatomically implausible features or textures, whereas the proposed method enforces consistency with the real-data distribution to keep perturbations plausible and visually similar to the input.Both approaches perturb the query image to flip the classification decision.Source: §1.1 · p. 3
  • Unlike prior approaches, the explanation function requires no retraining for new images at inference and uses classifier-consistency loss to make perturbations consistent with the classifier.Source: §1.1 · p. 3

1.2. Contributions

  • A cGAN generates progressively changing query-image perturbations until the classifier’s decision changes from negative to positive for a target class.Source: §1.2 · p. 3
  • A specialized reconstruction loss, informed by semantic segmentation and foreign-object detection, preserves anatomical shape and support devices across generated images.Source: §1.2 · p. 3
  • The authors evaluated the explanation function qualitatively and quantitatively while auditing a classifier trained on a chest X-ray dataset.Source: §1.2 · p. 3
  • They proposed explanation metrics based on the clinical definitions of cardiomegaly and pulmonary embolism, among the first such metrics for quantifying DNN explanations.Source: §1.2 · p. 3
  • The authors conducted a human-grounded study comparing counterfactual explanations for medical imaging using feedback from diagnostic radiology residents.Residents evaluated understandability, decision justification, visual quality, identity preservation, and overall helpfulness.Source: §1.2 · p. 4

2. Methodology

The methodology generates realistic, progressively modified images conditioned on desired classifier outcomes, using ordered-bin conditioning and consistency constraints. Context-aware reconstruction losses preserve input details and identity while encouraging smooth changes in classifier predictions.

2. Methodology

  • The explainer uses an image encoder, conditional GAN generator, and discriminator to produce a perturbed image with the desired target-class posterior while enforcing data, classifier, and reconstruction consistency.Its objective combines adversarial data-consistency, KL-divergence classifier-consistency, and reconstruction losses; reconstruction constraints preserve anatomical shape and foreign objects.Source: §2 · p. 4

2.1. Data consistency

  • An image encoder supplies the latent representation to a conditional GAN, which generates realistic-looking explanation images; the desired posterior is discretized into bins for conditioning.The desired posterior c ranges from 0 to 1, and changing it corresponds to changing the condition bin.Source: §2.1 · pp. 4–5

2.2. Classifier consistency

  • To encourage smooth prediction changes as the condition varies, the method treats condition bins as ordinal categories and checks that the desired bin exceeds all lower-index bins.The approach modifies the conditional discriminator loss to support ordinal multiclass regression.Source: §2.2 · p. 5
  • The condition-aware loss encourages the classifier’s score for the generated image to match the desired prediction distribution.It combines a conditional image-generation term with a KL-divergence term that influences the generator.Source: §2.2 · p. 6

2.3. Context-aware self consistency

  • The context-aware reconstruction loss (CARL) uses semantic segmentation and object detection information to preserve small or uncommon input details.Segmentation-based reconstruction penalizes differences in small regions, while detection information tracks input-image regions of interest.Source: §2.3 · p. 6
  • A separate reconstruction loss enforces self-reconstructed images to resemble the input and reverse perturbations to recover it.The classifier and the supporting segmentation and object-detection networks remain fixed during training.Source: §2.3 · pp. 6–7

3. Implementation and Evaluation

This section describes the MIMIC-CXR implementation and evaluation setup, reports classifier performance, and defines criteria for assessing counterfactual explanations. It evaluates clinically relevant image features using cardiothoracic ratio and healthy costophrenic-recess scoring, and illustrates progressive counterfactual changes for cardiomegaly and pleural effusion.

3.1. Dataset

  • Experiments used MIMIC-CXR, comprising 377K images from 65K patients with image-level labels for 14 observations.Source: §3.1 · p. 7

3.2. Implementation details

3.3. Evaluation

  • Evaluation used real-negative images with classifier scores below 0.2 and real-positive images above 0.8, generating counterfactuals for the negative group.Results were computed on the combined evaluation set.Source: §3.3 · p. 8
  • Evaluation metrics included FID for visual similarity, counterfactual validity for successful decision flips, and foreign object preservation for retaining detected objects.Source: §3.3 · p. 8
  • Clinical metrics translated cardiomegaly and pleural-effusion definitions into CXR measurements; CTR was used to quantify cardiomegaly.Source: §3.3 · p. 8
  • The cardiothoracic ratio (CTR) is cardiac diameter divided by the maximum internal diameter of the thoracic cavity; a CTR greater than 0.5 indicates cardiomegaly.Source: §3.3 · p. 8
  • CTR is calculated from segmented heart and lung regions, using the heart’s left-to-right extent and the lungs’ widest horizontal span.The heart diameter is measured from the lung centerline; the thoracic diameter is the horizontal distance between the widest points on the lung mask.Source: §3.3 · p. 8
  • For pleural effusion, the evaluation identifies costophrenic recesses and classifies them as healthy or blunt, with fluid accumulation associated with diaphragm flattening and angle blunting.A learned detector locates healthy or blunt recesses, and the healthy-recess score (SCP) is used as an evaluation metric.Source: §3.3 · p. 8
  • Counterfactual image sequences vary condition c from 0 to 1, with difference maps highlighting changed regions in cardiomegaly and pleural-effusion examples.The figure also shows classifier predictions, heart borders, and healthy or blunt costophrenic-recess detections.Source: Figure 4 · p. 9
  • Bounding-box annotation of the costophrenic region is simpler than expert-supervised CPA-angle marking, enabling a detector-based healthy-recess score (SCP).A blunt CPA indicates pleural effusion.Source: §3.3 · p. 9

4. Experiments and Results

This section evaluates counterfactual explanations for validity, visual quality, identity preservation, and changes in disease-specific clinical metrics, including comparisons with saliency maps and other methods. It also reports a diagnostic-radiology-resident study assessing understanding, helpfulness, decision justification, visual quality, and identity preservation.

4.1.1. Classifier consistency

  • Our method’s high counterfactual-validity score indicates successful decision flips, while CycleGAN flips decisions in only about 50% of cases.The paper also reports that classifier responses rise with the requested condition across input groups, consistent with the intended progressive behavior.Source: Table 1 · p. 10, §4.1.1 · p. 10

4.1.2. Visual quality

  • CycleGAN achieved the lowest FID, while our method had lower FID than xGEM and its explanations looked visually similar to query images.The authors characterize xGEM’s high FID as producing images too different from the query for deployment, whereas CycleGAN produced the most realistic explanations.Source: Table 1 · p. 10, §4.1.2 · pp. 10–11

4.1.3. Identity preservation

  • Table 2 reports the foreign-object preservation (FOP) score for the model with and without context-aware reconstruction loss (CARL); the score depends on the performance of the foreign-object detector.Source: Table 2 · p. 11

4.2. Comparison with saliency maps

  • The method supplements saliency maps with counterfactual images showing how relevant-region features must change to flip the decision; deletion evaluation removes top-ranked pixels to assess relevance.Saliency maps may highlight the same anatomical regions for multiple diagnoses, while the deletion experiment compares methods by progressively removing pixels ranked as important.Source: §4.2 · p. 11

4.3. Disease-specific evaluation

  • Figure 9 and paired tests showed statistically significant real-versus-counterfactual differences in CTR for cardiomegaly and SCP for PE.The authors evaluated cardiothoracic ratio (CTR) and score of the normal costophrenic recess (SCP) as disease-specific clinical metrics.Source: Figure 9 · p. 12, §4.3 · pp. 12–13

4.2. Comparison with saliency maps

  • Our method achieved the lowest AUDC, indicating the greatest classifier sensitivity in the salient regions it identified.The study treated a sharper decline in classification probability as evidence that removed pixels were important; a lower AUDC indicates that decline.Source: §4.2 · p. 12

4.4. Human evaluation

  • Only the counterfactual explanation significantly improved understanding of the classifier’s decision versus no explanation, and it outperformed CycleGAN on decision justification.Participants found evidence for the classifier’s decision more often with the counterfactual explanation than with CycleGAN.Source: §4.4 · p. 13
  • CycleGAN explanations scored better on visual quality and identity preservation, but were sometimes identical to the query image; participants found the counterfactual method most helpful overall.Participants rated the counterfactual method as most helpful for understanding the AI system’s assessment compared with the other explanation methods.Source: §4.4 · p. 13

4.4.1. Experiment Design

  • Twelve diagnostic radiology residents evaluated six randomly ordered CXRs under no-explanation, saliency-map, CycleGAN, and counterfactual conditions.The survey included diagnosis tasks before explanation conditions.Source: §4.4.1 · p. 13
  • The study compared understandability, decision justification, visual quality, identity preservation, and helpfulness across explanation conditions.Twelve diagnostic radiology residents participated; ratings used a five-point Likert scale except for helpfulness, which participants selected directly.Source: §4.4.1 · p. 14

4.4.2. Data analysis

  • For understandability and helpfulness, the authors used one-way ANOVA and Tukey HSD pairwise tests; they used paired t-tests to compare cycleGAN and counterfactual ratings on three other metrics.Each question received 72 responses, based on six chest x-rays and 12 participants.Source: §4.4.2 · p. 14

4.4.3. Results

  • Counterfactual explanations scored higher than the no-explanation baseline on understandability (p = 0.018); other explanation pairs did not differ significantly.The authors report that saliency maps and cycleGAN explanations did not significantly improve understandability over the baseline.Source: Figure 10 · p. 15, §4.4.3 · p. 15
  • Participants often linked understanding without an explanation to believing the classification was correct or assuming the AI used their reasoning; they found saliency maps lacked relevant regions.For cardiomegaly, participants noted that maps highlighted the heart but not its border or thoracic diameter.Source: §4.4.3 · p. 15
  • CycleGAN’s minimal perceptible changes impeded understanding (n=3), whereas 33 participants explicitly reported improved understanding with counterfactual progressions.Some participants said the progression helped them understand which features the AI was using.Source: §4.4.3 · p. 15
  • For classifier decision justification, the counterfactual explanation scored M=3.46 (SD=1.12), and participants found evidence for the predicted cardiomegaly class more frequently with counterfactual explanations than with cycleGAN.Source: §4.4.3 · p. 15
  • Participants noticed cardiomegaly-related changes, such as an enlarged heart, more in the explanation video than in cycleGAN (n=17), while cycleGAN changes were not perceptible (n=23).Participants reported changes not related to cardiomegaly, and the classifier may have associated such secondary information (“short-cuts”) with cardiomegaly.Source: §4.4.3 · pp. 15–16
  • Counterfactual explanations scored lower than cycleGAN by 0.31 on visual quality and 0.37 on identity preservation.Participants sometimes judged the counterfactual images nearly identical to the query, limiting the information they conveyed about the classifier’s decision.Source: §4.4.3 · p. 16
  • 57% of responses selected the counterfactual explanation as most helpful; E4 had the highest mean understandability, while its mean usefulness differed significantly from all other explanation conditions (p < 0.0001).Source: §4.4.3 · p. 16, Table 3 · p. 16

5. Discussion And Conclusion

The BlackBox Counterfactual Explainer generates image transformations that flip a classifier’s prediction, aiming to make medical image decisions more interpretable. In the reported evaluation, radiologists preferred its explanations for understandability, while the study also identified limitations and future directions.

Framework evaluation

  • The framework adhered to the three stated properties of valid counterfactual transformations: data consistency, classifier consistency, and self-consistency.It was proposed to explain a decision by gradually transforming an input image into a counterfactual that flips the classifier’s prediction.Source: §5 · p. 16

Comparison with xGEM and cycleGAN

  • Compared with the explainer, xGEM produced unrealistic images with a high FID score, while cycleGAN rarely flipped the classifier’s decision, with a CV score of ∼50%.The authors describe their method as producing visually appealing explanations while retaining patient-specific information.Source: §5 · pp. 16–17

Human evaluation

  • Radiologists found the counterfactual explanations most helpful for understanding the classifier’s decision and preferred them for understandability, despite reported visual inconsistencies.CycleGAN images were rated highly for quality and resemblance to the query, but were also seen as too similar to provide counterfactual reasoning.Source: §5 · p. 17

Clinical relevance of the explanations

  • Counterfactual changes associated with normal or abnormal predictions also tracked corresponding changes in disease-specific metrics, including CTR and SCP.The authors report that radiologists used CTR to diagnose cardiomegaly and present CTR calculations as a quantitative way to evaluate real and counterfactual images.Source: §5 · p. 17

Limitations

  • Generated counterfactuals may omit small details such as wires, although extended experiments found foreign objects such as pacemakers minimally important to classification.Source: §5 · p. 17
  • The foreign-object preservation score remained imperfect even with CARL, possibly because the object detector used to calculate it had limited capacity.Training a more accurate foreign-object detector was outside the study’s scope.Source: §5 · p. 17

Limitations and future work

  • At 256 × 256, counterfactual images were smaller than standard CXR and limited fine-detail evaluation; fewer conditions produced coarse transformations with abrupt changes, while a GPU-memory-fitting balanced batch resulted in stability.Source: §5 · p. 17

Conclusion

  • The model explained decisions for cardiomegaly and pleural effusions, with its explanations corroborated by an experienced radiology resident physician.The study frames counterfactual explanations as a way to audit whether a classifier uses radiographic features with clinical relevance.Source: §5 · p. 17

6. Supplementary Material … 6.3. Dataset

The supplementary material documents the human-evaluation questionnaire, summarizes notation, and describes the MIMIC-CXR dataset and label construction.

6.1. Human evaluation

  • The human evaluation asked 15 questions per chest x-ray, including diagnosis, agreement with the classifier, and understanding its assessment.Source: §6.1 · pp. 19–20
  • For saliency maps, cycleGAN, and the proposed explanation, questions assessed relevance to Cardiomegaly and whether the changes supported the classifier’s assessment.The proposed explanation questions also asked whether highlighted anatomical changes would change the Cardiomegaly assessment.Source: §6.1 · p. 20
  • Questions about cycleGAN and the proposed explanation assessed whether generated images looked like chest x-rays and preserved the subject’s appearance.Source: §6.1 · p. 20
  • The final question asked which explanation helped most, offering a heat map, a video of transformation from negative to positive, endpoint images, or none.Source: §6.1 · p. 20

6.2. Summarizing the notation

  • Table 4 summarizes the notation used in the manuscript.Source: §6.2 · p. 21

6.3. Dataset

  • MIMIC-CXR contains 473K chest x-rays and 206K reports from 63K patients; the study used frontal posteroanterior or anteroposterior views.Source: §6.3 · p. 21
  • Fourteen image-level radiographic labels were extracted from reports; the authors extracted MIMIC-CXR labels themselves because they had report access.The Stanford CheXpert labeler identifies positive, negative, or uncertain mentions and aggregates observations into labels (Irvin et al., 2019); its lexicon defines thoracic observations.Source: §6.3 · p. 21

6.4. Classification Model · 6.5. Explanation Function

The classification model is DenseNet-121, while the explanation method is a conditional GAN designed to explain black-box classifiers. The sections describe its architecture, training procedure, and selection of training conditions.

6.4. Classification Model

  • The classifier uses DenseNet-121 for diagnosis classification.The approach is intended to explain the model post hoc without access to its learned parameters or architectural details.Source: §6.4 · p. 21
  • The post-hoc approach does not require access to a model’s learned parameters or architecture and can explain any deep-learning neural network.Source: §6.4 · p. 21

6.5. Explanation Function

  • The explanation function is an encoder-equipped conditional GAN, with ResNet architectures for its encoder, generator, and discriminator.Source: §6.5 · p. 21
  • The encoder uses five ResBlocks with batch normalization and average-pool downsampling; the five-block generator uses conditional batch normalization and nearest-neighbor upsampling.Source: §6.5 · p. 21
  • The discriminator uses spectral normalization and downsamples after the second convolution in each ResBlock.Source: §6.5 · p. 21
  • Training used Adam (α = 0.0002, β1 = 0, β2 = 0.9) and updated the discriminator five times per generator update.Source: §6.5 · pp. 21–22
  • Experiments used N = 10 with batch size 32; the largest N was selected to create a class-balanced batch that fit GPU memory and resulted in stable cGAN training.Source: §6.5 · p. 22
  • Smaller δ increases the number of conditions, model complexity, training time, and batch-size needs, while GPU memory limits N; small N produces coarse transformations with abrupt image changes.Source: §6.5 · p. 22

6.6. Semantic Segmentation · 6.7. Object Detection

The merged sections describe a U-Net pipeline for lung and heart segmentation and object detectors for medical devices and costophrenic recesses. The detectors support context-aware reconstruction and evaluation, with the CP-recess detector reserved for evaluation.

6.6. Semantic Segmentation

  • A 2D U-Net segmented the lung and heart contours in CXRs using a multi-categorical cross-entropy loss over background, lung, and heart labels.Source: §6.6 · p. 22
  • The segmentation network was trained on 385 CXRs and corresponding masks from the JSRT and Montgomery datasets.Source: §6.6 · p. 23

6.7. Object Detection

  • For medical-device detection, reports supplied observation keywords, and 300 positive CXR images per observation were manually annotated with bounding boxes.The annotations marked foreign objects such as pacemakers, screws, and other hardware.Source: §6.7 · p. 23
  • A VGG-16-based fast Region-based CNN detector identified medical devices in CXRs and was used to enforce the context-aware reconstruction loss.The detector was trained on MIMIC-CXR.Source: §6.7 · p. 23
  • Separate detectors identified normal and abnormal CP recesses, labeled from reports mentioning “lungs are clear” and “blunting of costophrenic angle,” respectively.The study used 300 positive-mention images for each category.Source: §6.7 · p. 23
  • The CP-recess detector was used only for evaluation, not to train the explanation function; the Score for detecting a normal CP recess (SCP) was used as a metric.Source: §6.7 · p. 23
  • Instead of marking the CPA angle, the method annotates the CP region; CPA-angle marking on a CXR requires an expert to mark three points and calculate the angle and is prone to error.Source: §6.7 · p. 23

6.8. xGEM · 6.9. cycleGAN · 6.10. Extended results for identity preservation

The sections describe xGEM and cycleGAN comparators and extend identity-preservation evaluation. The reported tests found negligible prediction change after pacemaker removal and that counterfactuals remain similar to queries while differing in features sufficient to flip classification.

6.8. xGEM

  • xGEM iteratively traverses an input’s latent space to flip the classifier’s decision, but its VAE-based explanations can be unrealistic and over-smoothed.The explanation’s visual quality is limited by the generator’s expressiveness; vanilla VAE uses a Gaussian likelihood that is described as an unrealistic assumption for image data.Source: §6.8 · p. 23

6.9. cycleGAN

  • For each target disease, cycleGAN was trained on 2,000 positive-prediction images and an equal-sized set with strong negative predictions.The positive set had f(x) > 0.9 and the negative set had f(x) < 0.1.Source: §6.9 · p. 24

6.10. Extended results for identity preservation

  • The section motivates identity preservation by noting that a foreign object’s disappearance from an x-ray could be mistaken for the cause of a changed classification.The passage identifies FO as critical for patient identification in an x-ray.Source: §6.10 · p. 24

6.10.1. Ablation study over pacemaker

  • The average prediction change after pacemaker removal was 0.03, indicating that pacemakers did not influence cardiomegaly classification decisions in the tested subjects.The ablation included 300 subjects positively predicted for cardiomegaly who had a pacemaker; the pacemaker region was masked and in-filled with surrounding context.Source: §6.10.1 · p. 24

6.10.2. Latent space closeness (LCS)

  • Latent-space closeness (LSC) compares CARL explanations with those using simple ℓ1 reconstruction loss, scoring how often an explanation is closer to the query in feature-embedding distance.Similarity was calculated using Euclidean distance between embeddings from a pre-trained Inception v3 feature extractor.Source: §6.10.2 · p. 25
  • A high LSC score together with a high CV score indicates that query and counterfactual images are fundamentally the same, differing only in features sufficient to flip classification.The passage points to Table 6 for LSC results and Fig. 19 for CV results; it provides no specific score values here.Source: §6.10.2 · p. 25

6.11. Extended classifier consistency results · 6.12. Evaluating class discrimination · 6.13. Ablation Study

The extended results show that the counterfactual explanations flip classifier decisions, while the class-discrimination evaluation tests task specificity and ablations show the distinct roles of the model’s losses.

6.11. Extended classifier consistency results

  • A positive slope parallel to y = x shows the method flips decisions, producing high f(xc) from low f(x) and vice versa.The authors interpret this as representing the classifier’s decision-making process.Source: §6.11 · p. 26
  • For CycleGAN, explanations from inputs with f(x) ∈ [0.0, 0.4] remain below f(xc) = 0.5, failing to flip the decision and yielding a low CV score.Source: §6.11 · p. 26

6.12. Evaluating class discrimination

  • The class-discrimination evaluation tests cardiomegaly, pleural effusion, and edema by measuring flips in other classes; ideally, flips are highest for the target task and low elsewhere.The diagonal values also represent the counterfactual validity (CV) score.Source: §6.12 · p. 26

6.13. Ablation Study

  • The ablation assigns adversarial, KL, and CARL reconstruction losses to data consistency, classifier consistency, and context-aware self-consistency, respectively.Source: §6.13 · p. 26
  • Without adversarial loss (λ1 = 0), generated images have very high FID and low FOP because they look unlike real CXRs.Source: §6.13 · p. 26
  • Without KL classifier-consistency loss (λ2 = 0), CV is poor because explanations are generated without considering the classifier and fail to flip its decision.Source: §6.13 · p. 26
  • Without CARL loss (λ3 = 0), explanations no longer depict the query patient, omitting query-CXR findings and producing low FOP.Source: §6.13 · p. 26

6.14. Extended results for saliency maps

This section compares saliency maps from the counterfactual model with gradient-based methods using deletion evaluation. It reports that all methods localize important regions, while counterfactuals additionally show how relevant features should change to flip a decision.

6.14. Extended results for saliency maps

  • Because the model does not directly produce saliency maps, the authors approximate one with the absolute difference between explanations at normal and abnormal decision extremes.The extremes are defined as f(xc) < 0.2 for normal and f(xc) > 0.8 for abnormal; examples are shown in Fig. 8 and extended results in Fig. 21.Source: §6.14 · p. 26
  • For deletion evaluation (Petsiuk et al., 2018), the authors selectively inpainted pixels using their surroundings and calculated AUDC for 500 images.The metric tracks how target-class probability changes as important pixels are removed; the text refers to an unavailable table for these results.Source: §6.14 · p. 27
  • Low AUDC suggests all methods localize important classification regions, while counterfactual explanations also show which features should change to flip the decision.The authors caution that classifier behavior becomes inconsistent as deletion makes images unrealistic and visually unlike chest X-rays.Source: §6.14 · p. 27

6.15. Disease-specific evaluation

The evaluation uses clinical radiographic metrics to assess counterfactuals; for cardiomegaly, CTR differences between real and counterfactual images accompany classifier flips and distinguish real normal from abnormal groups.

6.15. Disease-specific evaluation

  • The evaluation compared real normal and abnormal groups with counterfactual groups using CTR and SCP, applying paired and independent two-sample t-tests.Counterfactual groups were defined by classifier-flipping images across the stated normal and abnormal prediction thresholds.Source: §6.15 · pp. 27–28
  • -0.03 (p < 0.0001) was the mean CTR difference for CTR(Xn) − CTR(Xp^c), supporting a statistically significant difference between normal and counterfactual abnormal images.The reported unpaired comparison also gives mean CTR(Xp) = 0.56; the supplied text does not provide the corresponding normal-group mean.Source: Table 8 · p. 28, §6.15 · p. 28
  • Small CTR changes in images initially predicted negative for cardiomegaly could flip the classifier, while significant CTR changes in real abnormal images led to a flip toward normal.The text relates the small changes to the small mean difference for normal images and their abnormal counterfactuals, and the larger changes to a large mean difference between abnormal and normal images.Source: §6.15 · p. 28

A. Input Image (Negative Cardiomegaly)Normal Abnormal · B. Input Image (Positive Cardiomegaly)

The examples compare counterfactual explanations for negative and positive cardiomegaly inputs, with a smaller CTR change for negative inputs. Separately, SCP scores distinguish normal from abnormal chest X-rays for pleural effusion.

B. Input Image (Positive Cardiomegaly)

  • For cardiomegaly, the model generates normal and abnormal explanation images and uses their pixel-wise difference as a saliency map.The yellow contour marks the heart boundary learned by a segmentation network; CTR denotes cardiothoracic ratio.Source: Figure 21 · p. 29

A. Input Image (Negative Cardiomegaly)Normal Abnormal

  • The cardiomegaly examples show a relatively smaller CTR change between real and counterfactual images for negatively classified inputs than for positively classified inputs.Figure 21 labels the negative-input examples in column A and the positive-input examples in column B.Source: Figure 21 · p. 29

B. Input Image (Positive Cardiomegaly)

  • Normal subjects had a mean SCP(Xn) of 0.69, versus 0.42 for abnormal patients (SCP(Xp),) in the pleural-effusion comparison.The authors report that SCP captured the difference between normal and abnormal chest X-rays.Source: B. Input Image (Positive Cardiomegaly) · p. 29

Related papers

Loading 2101.04230v3…