Pith. sign in

REVIEW 3 major objections 5 minor 46 references

Toxicity Assessment in Preclinical Histopathology via Class-Aware Mahalanobis Distance for Known and Novel Anomalies

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A single pixelwise AI system can segment nine known liver lesion types and, without any training on them, flag rare pathologies as out-of-distribution, achieving a 0.16% false negative rate on pathological tissue.

desk verdict The headline error rates are optimized on the test set, so don't quote them as deployment estimates; the dataset and adaptive-threshold idea are worth a look anyway. read the letter →

arxiv 2602.02124 v2 pith:MHOTHFL3 submitted 2026-02-02 cs.CV cs.AIcs.LG

Olga Graf , Dhrupal Patel , Peter Gro{ss} , Charlotte Lempp , Matthias Hein , Fabian Heinemann This is my paper
classification cs.CVcs.AIcs.LG
keywords histopathologyanomalydetectionout-of-distributionMahalanobisdistancewhole-slideimagessemanticsegmentationpreclinicaltoxicityscreeningclass-specificthresholds
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

The paper claims that a single pixelwise AI system can do two jobs at once in preclinical drug-safety screening: it can segment a mouse-liver whole-slide image into healthy tissue and nine common lesion types, and it can also flag tissue appearances that fit none of those classes — including rare pathologies it never saw during training. The authors' central methodological move is class-specific anomaly thresholds: instead of drawing one global line that separates 'known' from 'unknown' tissue, they draw a separate line for each predicted class, because the natural variability of healthy liver and each lesion type is very different. With that adaptation they report that only 0.16% of pathological tissue is called healthy and 0.35% of healthy tissue is called pathological, while held-out single-cell apoptosis and staining artifacts are detected as unknown. If these numbers hold in prospective use, the system could relieve the pathologist bottleneck in toxicity studies, quantify dose-dependent effects at pixel level, and catch subtle lesions early enough to inform go/no-go decisions.

What carries the argument

The load-bearing component is class-aware Mahalanobis distance: after fitting a shared covariance and class-specific means to the ℓ2-normalized pixel features of every known tissue class, the anomaly score of a pixel is the negative Mahalanobis distance from its features to the mean of the class that the segmentation head predicted. Instead of a single global threshold, each class receives its own threshold set to the (1−p) quantile of that class's score distribution, which is what lets the system confirm rare in-distribution lesions as known while still flagging genuinely novel tissue as out-of-distribution. The remainder of the pipeline — a frozen pre-trained ViT with low-rank-adapted atte

What would settle it

Rerun the pipeline with the threshold parameter fixed using only the training and validation sets, then evaluate on an independent set of whole-slide images; if the false negative rate for pathological tissue exceeds the reported 0.16%, the safety claim fails. Alternatively, hold out a third lesion type (e.g., a subtle degeneration) with no training data and count how many of its pixels are flagged as out-of-distribution rather than healthy; if most are called healthy, the 'rare pathologies without training data' claim does not generalise beyond apoptosis and artifacts.

Watch

Extended reading notes

Core claim

The central claim is that a Mahalanobis-distance anomaly score, computed per pixel from the features of a frozen pre-trained Vision Transformer adapted with low-rank updates, can serve simultaneously as a semantic segmentation of known tissue states and as an out-of-distribution detector for unknown ones. The authors' main departure from prior post-hoc OOD methods is to assign each predicted class its own acceptance threshold, defined as the (1−p) quantile of that class's score distribution, rather than thresholding one pooled score distribution. This matters because per-class score distributions in histology are strongly heterogeneous: a global threshold that keeps healthy tissue in-distrib

Load-bearing premise

The headline safety numbers are obtained by choosing the operating point p on the test set to minimize balanced error, so they are best-case values for this test set rather than unbiased estimates of a fixed, pre-specified operating point.

Editorial extensions

If this is right

  • A pathologist could receive one fused WSI map showing both the type of each known lesion and the locations that deserve extra scrutiny as potentially novel pathology.
  • In efficacy studies, non-target organs such as liver could be screened automatically for early toxicity signals, producing quantitative dose-response curves (e.g., vacuolated area per dose group) that mirror a pathologist's read.
  • Because the false negative rate is averaged per anomaly class, the metric itself enforces the safety-first requirement that no lesion type — however rare — may be silently missed.
  • The adaptive threshold selection can be applied on top of any post-hoc OOD score, so the principle is not tied to the Mahalanobis distance alone.
  • The framework can run as a secondary safety layer in exploratory toxicology studies, highlighting regions of interest and reducing the risk of missing subtle tissue alterations.

Reading between the lines

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

  • A natural next experiment the authors do not run: hold out a visually subtle third lesion class (rather than apoptosis and staining artifacts) to see whether the OOD recall remains near total; near-OOD lesions that resemble healthy cytoplasm are likely to be the hardest case.
  • The class-specific thresholding idea may transfer to other medical-imaging segmentation tasks where score distributions differ strongly by tissue compartment, such as cardiac MRI or dermatopathology, wherever a single global OOD threshold is known to be miscalibrated.
  • Because the reported operating point is selected on the test set, real deployment would need a pre-registered threshold protocol; without one, the safety margin a pathologist can rely on remains unknown.
  • If the dataset is made public, the same low-rank-adapted backbone could be benchmarked against newer histopathology foundation models, and the bottleneck would likely shift from architecture to annotation coverage.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes a pixelwise anomaly-detection framework for H&E-stained mouse liver whole-slide images intended for preclinical toxicity screening. It combines a frozen DINOv2 ViT adapted with LoRA and a linear segmentation head, then computes class-wise Mahalanobis scores and adaptive per-class thresholds to separate known (in-distribution) anomalies from novel (out-of-distribution) anomalies. The authors contribute a manually annotated dataset from 742 WSIs across 44 studies, hold out apoptosis and artifact as OOD classes, and report FNR = 0.16% and FPR = 0.35% at p = 0.996 for Adaptive-Maha+. They also present a toxicology-study demonstration showing dose-dependent cytoplasmic vacuolation.

Significance. If the reported numbers held, this would be a practically important step: a single pixelwise map that segments nine lesion types and flags unseen morphologies, with near-zero missed pathological tissue, would be valuable for pathology workflow. The paper has strengths: a substantial annotated histopathology dataset, a sensible extension of Mahalanobis OOD detection to per-class thresholds, and a comprehensive comparison to seven post-hoc OOD methods under two thresholding strategies. However, the central quantitative claims are presently not supported because the operating point p is chosen on the test set and because the FNR metric is a per-class average under an 'any anomaly' definition. The value of the contribution is conditional on a corrected evaluation protocol.

major comments (3)
  1. [Results, 'Performance on the test set'] The reported FNR=0.16% and FPR=0.35% are not estimates at a pre-specified operating point. The text states: for each method and threshold selection strategy, the authors select p from the evaluated range that minimizes the balanced error rate (BER), defined as the mean of FNR and FPR, on the test set. Thus the headline numbers are the minimum of a BER curve over p in [0.95,1.00) using test labels. Table 1 reports each method at its own test-set-optimal p, so the absolute rates and the cross-method ranking are optimized on the test set; the 5-seed standard errors do not include the variability of p selection. This directly undermines the abstract's claim that the system achieves these rates. Please re-select p on the training/validation set only, or report full BER/FNR/FPR curves with a fixed selection rule, and restate the headline numbers. The abstract's description of 'optimizing the f
  2. [Results, 'Unknown anomaly detection'; Abstract] The FNR definition counts any anomaly-as-anomaly prediction as a true positive and averages per-class FN rates: FNR = (1/(K+1)) * sum_i FN_i/(TP_i+FN_i). Therefore the abstract's 'only 0.16% of pathological tissue classified as healthy' is misleading: it is a mean over classes under a lenient 'anything but healthy' criterion, not the global fraction of pathological pixels, and cross-type confusions (e.g., necrosis predicted as inflammation) are treated as correct. The paper does disclose this, but the headline wording does not, and the safety-first 'never overlooking a lesion' claim relies on the lenient definition. Please report strict per-class FNR where a correct prediction requires the exact class, and the overall pixel-level healthy-vs-pathology rate, and adjust the abstract accordingly.
  3. [Data splits; 'Detection of OOD anomalies on the whole-slide image'] The general claim that rare pathologies are detected 'without training data' is supported by only two held-out classes, apoptosis and artifact, both from the same dataset and scanner. While apoptosis is a sensible near-OOD stress test, two classes do not license the broad conclusion about unseen morphologies. Please either restrict the claim to these two types or validate on additional held-out classes or studies. This is a load-bearing qualification for the abstract's 'rare pathologies' claim.
minor comments (5)
  1. [Abstract] The phrase 'Optimizing the false positive rate subject to a predefined constraint on the false negative rate' does not match the BER-minimization procedure described in the Results section; revise for consistency.
  2. [Appendix A2] The heading 'Averaging across spacial shifts' contains a typo: 'spacial' should be 'spatial'.
  3. [Table 1 and Table 2] The sentence 'Values marked with indicate improvement' appears to have a missing symbol; please fix the rendering.
  4. [Figure 10 / Appendix A3] The confusion matrices in Fig. 10 are computed on the joint training and validation set, not the test set. This should be stated explicitly when the figure is referenced in the main text.
  5. [Figure 11 / Appendix A4] The t-SNE visualization is useful but is discussed only qualitatively. Consider adding an objective separability measure to support the claim that the OOD class occupies a near-OOD region.

Circularity Check

2 steps flagged · score 6.0 of 10

Headline FNR/FPR are obtained by selecting p on the test set to minimize BER, so the reported 0.16%/0.35% are optimized, not independently predicted, values.

  1. fitted input called prediction [Results and Discussion, 'Performance on the test set' (after Table 1); Table 1 caption]
    "For each combination of method and threshold selection strategy, we select p from the evaluated range that minimizes the balanced error rate (BER), defined as the mean of FNR and FPR, on the test set. ... We observe that p=0.996 yields the best BER for our approach (Adaptive-Maha+). ... The best FNR of 0.16% is achieved by Maha+ under the adaptive threshold selection strategy. Each combination of method and threshold selection strategy is reported at its best achievable performance and the corresponding p is given."

    The abstract's central quantitative claim (0.16% pathological tissue classified as healthy and 0.35% healthy tissue classified as pathological) is not the performance of a pre-specified operating point. The free operating parameter p is chosen on the test set by minimizing BER; for Adaptive-Maha+ this selects p=0.996, and the reported FNR/FPR are the values at that test-set-optimal p. Thus the headline error rates are the minimum of a BER curve over the evaluated p-range, i.e. optimistic fitted quantities rather than independent estimates. Moreover, Table 1 reports every method 'at its best achievable performance' with its own best p, so the cross-method ranking is confounded by per-method test-set selection; the +/- standard errors over 5 seeds do not include the variance of the p-selecti

  2. other [Abstract]
    "Optimizing the false positive rate subject to a predefined constraint on the false negative rate yields only 0.16% of pathological tissue classified as healthy and 0.35% of healthy tissue classified as pathological."

    The abstract describes the threshold optimization as constraining FNR and then optimizing FPR, but the Results section defines the actual procedure as selecting p to minimize BER (mean of FNR and FPR) on the test set. There is no predefined FNR constraint in the reported protocol. This mismatch matters because the quoted numbers are the output of that test-set optimization, so presenting them as the consequence of a safety-first constraint overstates what the evaluation protocol supports. The claim does not add independent evidence beyond the test-set-fitted p, and it is contradicted by the paper's own methodological description.

full rationale

The core technical derivation is not circular: the DINOv2/LoRA segmentation features, the class-conditional Mahalanobis score, and the adaptive per-class threshold quantiles are all computed from training/validation data and are externally meaningful; no self-citation chain or uniqueness argument is used to force the result. The circularity is confined to the evaluation protocol. The single scalar p controlling all class thresholds is selected on the test set to minimize BER, and the abstract's 0.16%/0.35% as well as Table 1's ranking are reported at each method's own test-set-optimal p. The Methods sentence 'The threshold selection was performed on joint training and validation set' does not remove this issue, because the Results explicitly state that the p-range is evaluated on the test set and the best p is chosen by test-set BER. Since p is a one-dimensional operating point with a direct monotonically related effect on the reported FNR/FPR, reporting the minimum of the BER curve as the achieved rate is a fitted-input-called-prediction step, not an unbiased estimate of deployment performance at a fixed threshold. I therefore score 6: partial circularity in the central quantitative claim, while the underlying method itself retains independent content. If the paper were reframed as a method proposal with p selection acknowledged as a benchmark selection, the circularity would drop substantially; as written, the headline number is statistically forced by the test-set fit.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The central claim relies on standard statistical assumptions of the Mahalanobis method, on transferability of DINOv2 features, on the representativeness of two OOD classes, on partial annotations being sufficient ground truth, and on the test set being used only for final evaluation. The last assumption is violated by the test-set-based selection of p, which is the main reason the headline error rates cannot be taken at face value.

free parameters (6)
  • p (target ID acceptance fraction) = 0.996 (chosen on test set to minimize BER)
    Searched over [0.95,1.00) on the test set, not validation; directly controls FNR/FPR operating point.
  • LoRA rank r = 3
    Hand-chosen; no ablation reported; controls capacity of the adaptation.
  • Class-specific minimum annotated crop fractions = 1% (most classes), 0.5% (vacuolation, mitosis), >=1 px (microsteatosis)
    Data-sampling filters chosen by hand; affect which pixels the model sees and therefore segmentation quality.
  • Spatial shift stride / number of shifts = 84 px / 36 shifts
    Inference-time averaging configuration chosen by hand; affects robustness and computational cost.
  • Class-specific thresholds tau_i,p = not reported; (1-p)-quantiles of per-class score distributions
    The core operating thresholds; estimated on train+val, but the defining p is selected on the test set.
  • Class-conditional feature means and shared covariance = estimated from train+val features
    Central to the Mahalanobis distance; estimated from data and not independently validated.
assumptions (5)
  • domain assumption Class-conditional feature distributions are multivariate Gaussian with a shared covariance matrix.
    Appendix A5 assumes p(h|y=c_i)=N(h; mu_c_i, Sigma). If features are not Gaussian, the Mahalanobis thresholds are miscalibrated.
  • domain assumption DINOv2 frozen features transfer to H&E mouse liver at 0.442 um/px.
    The whole approach rests on foundation-model features being semantically meaningful for this domain; no feature-quality check beyond segmentation IoU.
  • domain assumption Apoptosis and artifact are representative of all novel toxicologic anomalies.
    Only two OOD categories are evaluated; the claim of detecting rare pathologies without training data generalizes beyond these.
  • domain assumption Partial annotation is adequate for evaluation.
    Only selected regions are annotated; unlabeled pathology in WSIs (e.g., Fig. 7E) is not counted in error rates, biasing FNR/FPR estimates.
  • ad hoc to paper Selecting p on the test set does not invalidate reported performance.
    The paper states p is selected on the test set to minimize BER; the reported numbers are therefore fitted to the evaluation data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Toxicity Assessment in Preclinical Histopathology via Class-Aware Mahalanobis Distance for Known and Novel Anomalies." pith.science (2026). https://pith.science/paper/MHOTHFL3

@misc{pith2026260202124,
  author       = {Pith},
  title        = {Pith review of: Toxicity Assessment in Preclinical Histopathology via Class-Aware Mahalanobis Distance for Known and Novel Anomalies},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MHOTHFL3}},
  note         = {Machine review of arXiv:2602.02124}
}
read the original abstract

Drug-induced toxicity is a leading cause of preclinical and early-clinical failure, making early detection critical. Histopathology is the gold standard for toxicity assessment but relies on expert pathologists, creating a bottleneck for large-scale screening. We introduce an AI-based anomaly detection framework for whole-slide images (WSIs) of rodent liver that identifies healthy tissue and known pathologies (anomalies) and flags samples without training data as out-of-distribution (OOD). We evaluate OOD detection on two held-out categories: apoptosis (single-cell, near-OOD) and staining/processing artifacts (heterogeneous, far-OOD). We build a novel pixelwise-annotated dataset and fine-tune a pre-trained Vision Transformer (DINOv2) via Low-Rank Adaptation (LoRA) for segmentation, then use the Mahalanobis distance for OOD detection with class-specific thresholds. Optimizing the false positive rate subject to a predefined constraint on the false negative rate yields only 0.16% of pathological tissue classified as healthy and 0.35% of healthy tissue classified as pathological. Our false negative rate does not penalise cross-type errors, reflecting the safety-first objective of never overlooking a lesion; under the stricter correct-class criterion our method assigns 93.93% of ID and 89.38% of OOD findings to their own class. The study demonstrates technical feasibility of pixel-level anomaly detection for mouse liver histopathology, indicating possible applications in improving preclinical workflows and drug development efficiency.

Figures

Figures reproduced from arXiv: 2602.02124 by the authors.

Figure 1
Figure 1. Two-step training approach of the system that performs semantic segmentation of tissue states with available training data (in-distribution (ID) data; i.e., healthy tissue and a set of common pathologies); additionally, it can detect tissue states without training examples as out-of-distribution (OOD) anomalies (i.e., rare pathologies, or artifacts). A. In Step 1, the system is trained as a conventional segmentation… view at source ↗
Figure 2
Figure 2. Inference workflow for segmentation and anomaly detection. WSIs (H&E-stained mouse liver) are divided into patches and processed by an encoder to extract spatially resolved feature representations (h). These features are passed to a classifier to predict the most likely class (ci), and the outputs are averaged across spatial shifts for increased robustness. For each predicted class, the Mahalanobis distance (sMaha+)… view at source ↗
Figure 3
Figure 3. Schematic representation of spatial averaging procedure applied to improve robustness during inference and mitigate boundary artifacts. Augmentation and averaging using spatial shifts To improve robustness and generalization, we apply spatial shifting strategy. Unlike in classical semantic segmentation tasks that operate on a full image, WSIs must be divided into smaller tiles for computational feasibility. Applying… view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Extended confusion matrix for unified segmentation and anomaly detection. To evaluate the effectiveness of anomaly detection – both known and unknown – we use FNR and FPR. They allow to assess the percentage of all missed anomalies and false alarms, respectively, which…
Figure 5
Figure 5. Figure 5: Distribution of the mean anomaly score per tile for healthy (no damage model was induced) tissue samples. The samples with the highest scores are the most normal. The outliers with lower scores (i.e., rated as more anomalous) are samples with low glycogen storage as we…
Figure 6
Figure 6. Figure 6: Anomaly score histograms for different anomaly classes and the normal (healthy tissue) class, calculated on combined training and validation set. Using a fixed threshold (black dashed line) for all classes to determine OOD samples would lead to many errors for a lot of…
Figure 7
Figure 7. Figure 7: Whole-slide image analysis of mouse liver stained in H&E from a CCl4-induced disease model to study metabolic dysfunction-associated steatohepatitis (MASH). A. Example areas of WSI showing the original image and the ground truth annotated by a pathologist. B. The same …
Figure 8
Figure 8. Figure 8: Detection of adverse drug reactions in a preclinical toxicological liver study. Our approach (Adaptive-Maha+) disentangles dose-dependent toxicologic effects from vehicle-related and incidental anomalies at pixel level. A. Percentage of liver area occupied by different…
Figure 9
Figure 9. Figure 9: Performance of OOD detection methods under Standard (A) and Adaptive (B) threshold selection across various target percentages of predicted ID samples. Balanced error rate (BER), false negative rate (FNR) and false positive rate (FPR) are evaluated on the test set at d…
Figure 10
Figure 10. Figure 10: Extended confusion matrices for unified segmentation and anomaly detection, evaluated with Standard-Maha+ (p = 0.980) and Adaptive-Maha+ (p = 0.996) on joint training and validation set. A. Standard threshold selection. A substantial fraction of ID anomalies is miscla…
Figure 11
Figure 11. Figure 11: t-SNE visualization of feature representations of pixels on the test set. Overall, t-SNE illustrates separability for major tissue patterns, while rare or visually ambiguous classes exhibit overlap. The unknown class (Apoptosis/Artifact) occupies a near-OOD region of …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 1 canonical work pages

  1. [1]

    W., Blanckley, A., Boldon, H

    Scannell, J. W., Blanckley, A., Boldon, H. & Warrington, B. Diagnosing the decline in pharmaceutical R&D efficiency. Nat. reviews Drug discovery11, 191–200 (2012)

    2012

  2. [2]

    J.et al.An analysis of the attrition of drug candidates from four major pharmaceutical companies.Nat

    Waring, M. J.et al.An analysis of the attrition of drug candidates from four major pharmaceutical companies.Nat. reviews Drug discovery14, 475–486 (2015)

    2015

  3. [3]

    Everitt, J. I. The future of preclinical animal models in pharmaceutical discovery and development: a need to bring in cerebro to the in vivo discussions.Toxicol. Pathol.43, 70–77 (2015)

    2015

  4. [4]

    Zynger, D. L. & Pernick, N. Understanding the pathology job market: an analysis of 2330 pathology job advertisements from 2013 through 2017.Arch. pathology & laboratory medicine143, 9–10 (2019)

    2013

  5. [5]

    Reports12, 19236 (2022)

    Heinemann, F.et al.Deep learning-based quantification of NAFLD/NASH progression in human liver biopsies.Sci. Reports12, 19236 (2022)

    2022

  6. [6]

    medicine25, 1301–1309 (2019)

    Campanella, G.et al.Clinical-grade computational pathology using weakly supervised deep learning on whole slide images.Nat. medicine25, 1301–1309 (2019)

    2019

  7. [7]

    pathology 38, 5S–81S (2010)

    Thoolen, B.et al.Proliferative and nonproliferative lesions of the rat and mouse hepatobiliary system.Toxicol. pathology 38, 5S–81S (2010)

    2010

  8. [8]

    & Monticello, T

    Ying, X. & Monticello, T. M. Modern imaging technologies in toxicologic pathology: An overview.Toxicol. pathology34, 815–826 (2006)

    2006

Show all 46 references
  1. [9]

    C., Shawe-Taylor, J., Smola, A

    Schölkopf, B., Platt, J. C., Shawe-Taylor, J., Smola, A. J. & Williamson, R. C. Estimating the support of a high-dimensional distribution.Neural Comput.13, 1443–1471 (2001)

    2001

  2. [10]

    & Gimpel, K

    Hendrycks, D. & Gimpel, K. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations (ICLR)(2017)

    2017

  3. [11]

    InInternational Conference on Machine Learning (ICML)(2022)

    Hendrycks, D.et al.Scaling out-of-distribution detection for real-world settings. InInternational Conference on Machine Learning (ICML)(2022)

    2022

  4. [12]

    Liu, W., Wang, X., Owens, J. & Li, Y . Energy-based out-of-distribution detection. InAdvances in Neural Information Processing Systems (NeurIPS)(2020). 13.Mahalanobis, P. On the generalised distance in statistics.Proc. Natl. Inst. Sci. India2, 49–55 (1936)

    2020

  5. [14]

    & Shin, J

    Lee, K., Lee, H., Lee, K. & Shin, J. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. InProceedings of the 32nd Conference on Neural Information Processing Systems (NeurIPS 2018)(2018)

    2018

  6. [15]

    Sun, Y ., Guo, C. & Li, Y . React: Out-of-distribution detection with rectified activations. InAdvances in Neural Information Processing Systems (NeurIPS)(2021)

    2021

  7. [16]

    & Srikant, R

    Liang, S., Li, Y . & Srikant, R. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations (ICLR)(2018)

    2018

  8. [17]

    & Dietterich, T

    Hendrycks, D., Mazeika, M. & Dietterich, T. G. Deep anomaly detection with outlier exposure. InInternational Conference on Learning Representations (ICLR)(2019)

    2019

  9. [18]

    Sun, Y ., Ming, Y ., Zhu, X. & Li, Y . Out-of-distribution detection with deep nearest neighbors. InInternational Conference on Machine Learning (ICML)(2022)

    2022

  10. [19]

    InProceedings of the European Conference on Computer Vision (ECCV), 550–564 (2018)

    Vyas, A.et al.Out-of-distribution detection using an ensemble of self supervised leave-out classifiers. InProceedings of the European Conference on Computer Vision (ECCV), 550–564 (2018)

    2018

  11. [20]

    & Steger, C

    Bergmann, P., Löwe, S., Fauser, M., Sattlegger, D. & Steger, C. Improving unsupervised defect segmentation by applying structural similarity to autoencoders. InProceedings of the 14th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and A...

    2019

  12. [21]

    InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)(2019)

    Gong, D.et al.Memorizing normality to detect anomaly: Memory-augmented deep autoencoder for unsupervised anomaly detection. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)(2019)

    2019

  13. [22]

    Kingma, D. P. & Welling, M. Auto-encoding variational Bayes. InInternational Conference on Learning Representations (ICLR)(2014)

    2014

  14. [23]

    & Uehara, K

    Sato, K., Hama, K., Matsubara, T. & Uehara, K. Predictable uncertainty-aware unsupervised deep anomaly segmentation. InInternational Joint Conference on Neural Networks (IJCNN)(2019). 12/17

    2019

  15. [24]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)(2020)

    Liu, W.et al.Towards visually explaining variational autoencoders. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)(2020)

    2020

  16. [25]

    & Adeli, E

    Sabokrou, M., Khalooei, M., Fathy, M. & Adeli, E. Adversarially learned one-class classifier for novelty detection. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)(2018)

    2018

  17. [26]

    Pidhorskyi, S., Almohsen, R., Adjeroh, D. A. & Doretto, G. Generative probabilistic novelty detection with adversarial autoencoders. InAdvances in Neural Information Processing Systems (NeurIPS)(2018)

    2018

  18. [27]

    & Breckon, T

    Akcay, S., Atapour-Abarghouei, A. & Breckon, T. P. Ganomaly: Semi-supervised anomaly detection via adversarial training. InAsian Conference on Computer Vision (ACCV)(2018). 28.Liu, J.et al.A survey on diffusion models for anomaly detection.arXiv preprint arXiv:2501.11430(2025)

    2018 arXiv

  19. [29]

    & Xiang, B

    Perera, P., Nallapati, R. & Xiang, B. OCGAN: One-class novelty detection using GANs with constrained latent representa- tions. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)(2019)

    2019

  20. [30]

    & Guo, Y

    Zhao, W., Li, J., Dong, X., Xiang, Y . & Guo, Y . Segment every out-of-distribution object. In2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 3910–3920, DOI: 10.1109/CVPR52733.2024.00375 (2024)

    2024

  21. [31]

    In European Conference on Computer Vision (ECCV), 246–263 (Springer, 2022)

    Tian, Y .et al.Pixel-wise energy-biased abstention learning for anomaly segmentation on complex urban driving scenes. In European Conference on Computer Vision (ECCV), 246–263 (Springer, 2022)

    2022

  22. [32]

    InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 1151–1161 (2023)

    Liu, Y .et al.Residual pattern learning for pixel-wise out-of-distribution detection in semantic segmentation. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 1151–1161 (2023)

    2023

  23. [33]

    & Gholipour, A

    Karimi, D. & Gholipour, A. Improving calibration and out-of-distribution detection in deep models for medical image segmentation.IEEE Transactions on Artif. Intell.4, 383–397, DOI: 10.1109/TAI.2022.3159510 (2023)

    2022

  24. [34]

    Dippel, J.et al.AI-based anomaly detection for clinical-grade histopathological diagnostics.NEJM AI1, DOI: 10.1056/ AIoa2400468 (2024)

    2024

  25. [35]

    & Vercauteren, T

    Wang, J., Wang, Z., MacCormac, O., Shapey, J. & Vercauteren, T. OOD-SEG: Exploiting out-of-distribution detection techniques for learning image segmentation from sparse multi-class positive-only annotations. InarXiv:2411.09553(2025)

    2025 arXiv

  26. [36]

    & Heinemann, F

    Zingman, I., Stierstorfer, B., Lempp, C. & Heinemann, F. Learning image representations for anomaly detection: application to discovery of histological alterations in drug development.Med. Image Analysis92, 103067 (2024)

    2024

  27. [37]

    38.Hu, E

    Oquab, M.et al.DINOv2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193(2023). 38.Hu, E. J.et al.Lora: Low-rank adaptation of large language models.ICLR1, 3 (2022)

    2023 arXiv

  28. [39]

    J.et al.Towards a general-purpose foundation model for computational pathology.Nat

    Chen, R. J.et al.Towards a general-purpose foundation model for computational pathology.Nat. medicine30, 850–862 (2024)

    2024

  29. [40]

    medicine30, 2924–2935 (2024)

    V orontsov, E.et al.A foundation model for clinical-grade computational pathology and rare cancers detection.Nat. medicine30, 2924–2935 (2024)

    2024

  30. [41]

    & Hein, M

    Bitterwolf, J., Müller, M. & Hein, M. In or out? Fixing Imagenet out-of-distribution detection evaluation. InProceedings of the 40th International Conference on Machine Learning(2023)

    2023

  31. [42]

    & Hein, M

    Müller, M. & Hein, M. Mahalanobis++: Feature normalization as the missing ingredient for reliable OOD detection. In Proceedings of the 42nd International Conference on Machine Learning(2025). 43.Siméoni, O.et al.DINOv3.arXiv preprint arXiv:2508.10104(2025)

    2025 arXiv

  32. [44]

    neural information processing systems32(2019)

    Paszke, A.et al.Pytorch: An imperative style, high-performance deep learning library.Adv. neural information processing systems32(2019)

    2019

  33. [45]

    P., Rees, S., Kalindjian, S

    Hughes, J. P., Rees, S., Kalindjian, S. B. & Philpott, K. L. Principles of early drug discovery.Br. journal pharmacology 162, 1239–1249 (2011)

    2011

  34. [46]

    A., Sagartz, J

    Kramer, J. A., Sagartz, J. E. & Morris, D. L. The application of discovery toxicology and pathology towards the design of safer pharmaceutical lead candidates.Nat. Rev. Drug Discov.6, 636–649 (2007)

    2007

  35. [47]

    & van Diest, P

    Thoolen, B., Bradley, A., Stathonikos, N. & van Diest, P. J. Toxicologic pathology forum*: Opinion on the future of histopathology using whole slide images in toxicologic pathology of preclinical studies and its successful implementation in compliance with good laboratory prac...

    2025

  36. [48]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020)

    Dosovitskiy, A. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929(2020). 13/17

    2010 arXiv

  37. [49]

    & Farahmand, A

    Hao, Z., AghaKouchak, A., Nakhjiri, N. & Farahmand, A. Global integrated drought monitoring and prediction system (GIDMaPS) data sets. http://dx.doi.org/10.6084/m9.figshare.853801 (2014)

    2014 doi

  38. [50]

    Brereton, R. G. & Lloyd, G. R. Re-evaluating the role of the Mahalanobis distance measure.J. Chemom.30, 134–143 (2016). Acknowledgements Olga Graf is supported by the AI & Data Science Fellowship Post Doc program from Boehringer Ingelheim with the University of Tübingen, Germa...

    2016

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.