A Computer Vision Approach to Combat Lyme Disease

Elham Dolatabadi; Laurent Moreno; Samir Patel; Sina Akbarian; Tania Cawston; Vanessa Allen

REVIEW 4 major objections 7 minor 34 references

A Computer Vision Approach to Combat Lyme Disease

T0 review · 4 major / 7 minor · reviewed 2026-08-27 · deepseek-v4-flash

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

Pith's one-line read A compact convolutional neural network tells blacklegged ticks from other tick species in phone photos with 92% accuracy, a step toward timely Lyme prophylaxis.

desk verdict A useful new tick-image dataset and a plausible triage classifier, but the 92% accuracy claim hinges on an under-specified train/test split that could let the model recognize individuals rather than species. read the letter →

arxiv 2009.11931 v1 pith:DVYJDLG4 submitted 2020-09-24 cs.CV cs.LGeess.IV

Sina Akbarian , Tania Cawston , Laurent Moreno , Samir Patel , Vanessa Allen , Elham Dolatabadi This is my paper · ORCID

classification cs.CVcs.LGeess.IV

keywords Lymediseaseblackleggedtickidentificationconvolutionalneuralnetworkattentiontransferlabelsmoothingknowledgedistillationcomputervision

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

Lyme disease must be treated within 72 hours of a tick bite, but laboratory identification of the tick usually cannot meet that window. The paper sets out to show that a compact convolutional neural network can classify blacklegged ticks from ordinary phone photos quickly enough to inform the prophylaxis decision. It trains a small student network with attention transfer and label smoothing regularization, using a large pre-trained network as teacher. On a public health laboratory dataset of 12,588 images, the best model reaches 92% test accuracy. If the result holds on unseen ticks, a mobile or web tool could pair the species verdict with exposure geography to guide timely clinical action.

What carries the argument

The load-bearing mechanism is teacher-student knowledge transfer: an Inception-ResNet pre-trained on ImageNet acts as teacher, and a 13-layer lighter CNN with 5,350,633 trainable parameters acts as student. The teacher's spatial attention map is built as $Q = \sum_{i=1}^{C} |A_i|$ for the last layer's 3D tensor $A$, and an $\ell^2$-normalized attention-transfer loss pushes the student's pre-final-layer attention map to match. Label smoothing regularization then replaces one-hot labels with soft labels: a first student outputs temperature-softened class probabilities for correctly classified images and a constant 0.6 for misclassified ones, and a second student is trained on a weighted combination of cross-entropy and attention loss. The combination lets a small network imitate a much larger one, which the authors argue explains both the accuracy and the suitability for mobile deployment.

What would settle it

Re-run the evaluation with a split by tick identity, so both photos of each tick stay on the same side of the boundary, and compare accuracy with the reported 92%; a large drop would indicate the model memorized individual ticks. A complementary check is testing on phone photos of ticks from regions or seasons not represented in the training set.

Watch

Extended reading notes

Core claim

The central claim is that knowledge transfer from a large image-classification teacher network to a small student CNN, using spatial attention transfer plus label smoothing, yields a blacklegged-tick classifier that is both accurate (92% test accuracy, 97.32% ROC-AUC) and compact (5.35 million trainable parameters) enough to run on a smartphone. The student learns to reproduce the teacher's spatial attention map, computed by summing absolute activations across channels, and its class probabilities are smoothed by soft labels generated by a first student. The authors report that this AT+LSR student outperforms training from scratch and a large 53-million-parameter network, while fine-tuning an ImageNet-initialized network with most layers frozen collapses to 42% accuracy. The concrete discovery is that, on this dataset, a deliberately small network can capture the visual distinction between blacklegged and non-blacklegged ticks when guided by attention transfer and label smoothing.

Load-bearing premise

The 92% accuracy figure assumes the train/test split separates whole ticks rather than individual photos, because each tick contributes dorsal and ventral images, and the paper does not state the unit of the split.

Editorial extensions

If this is right

A phone-deployable classifier at 92% accuracy could give patients a same-day species identification, preserving the 72-hour antibiotic window that laboratory workflows often miss.
Pairing the classification with geolocation, as the paper's web application does, would let the tool estimate Lyme risk from the overlap of the identified species and known tick distribution.
The result suggests that attention transfer and label smoothing can close the accuracy gap between small and large networks on small medical-image datasets, pointing away from simply using bigger models.
If deployed publicly, uploaded photos could double as tick-surveillance data, helping track where blacklegged ticks are expanding.
The 42% accuracy of the frozen-layers Inception-Resnet indicates that fine-tuning only the final layers is not a viable strategy on this data, a caution for similar projects.

Reading between the lines

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

The split-unit ambiguity is the main threat to the accuracy claim; if the model was tested on photos rather than ticks, its true generalization to new ticks could be lower, and a per-tick split would settle it.
The same teacher-student recipe could transfer to other small-image classification problems in infectious disease diagnostics, such as mosquito species or rash photos, where labeled datasets are similarly scarce.
The 0.6 constant for misclassified images and the temperature T=5 are ad hoc; ablating them would reveal how much of the gain comes from label smoothing versus attention transfer.
A saliency-map analysis could show whether the model keys on species-specific markings or on body shape (engorged versus unfed), which would matter for the app's reliability across tick life stages.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Request a human review

A listed scientist reviews the paper for a fee and the review publishes here regardless of verdict. See the reviewers or get listed.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

You should know this paper is a mixed bag. It presents a new dataset of ~6,300 ticks photographed under fairly controlled conditions and applies standard transfer learning plus teacher-student attention transfer to classify blacklegged vs. other ticks. The lightweight student network reaching 92% accuracy, if it holds, is a nice practical result. The problem (Lyme disease prophylaxis within 72 hours) is real, and the authors have engaged with clinicians and built an internal web tool for validation.

What's genuinely useful: the dataset itself, the careful data collection protocol (phone images, dorsal/ventral pairs, engorged/unfed states), and the demonstration that a 5.3M-parameter student can match a 53M teacher. That is a legitimate contribution, even without new algorithmic ideas.

The soft spots are real but mostly fixable. The biggest one is that the train/test split is not described at the unit level. Each tick has dorsal and ventral phone photos, plus some high-res microscope shots. If the split is by image rather than by tick, the same tick can appear in both train and test, and the model can learn individual-level cues (size, markings, lighting of that tick's session). That would inflate the reported 92% accuracy and would not reflect species generalization. The authors say 'without any overlap' but never say whether overlap is defined on ticks or images. This needs to be clarified and re-evaluated if the split was image-level.

Other issues: The text says 12,588 images were captured, but the split sums to 13,588. The 1,000 high-res images explain the difference, but the text doesn't say that explicitly. The LSR recipe is non-standard: using a student network to generate soft labels, then replacing probabilities for misclassified images with a fixed 0.6, and the loss weights are written as denominators, which is odd. Also there is no comparison to any prior tick-classification work; the related work is generic deep learning. No code or data are released, so independent verification isn't possible.

None of this is fatal. The central claim—that a compact CNN can distinguish blacklegged ticks from other common species on this dataset—is plausible and probably true even with a tick-level split, though the accuracy number may drop. The paper deserves to be sent to reviewers, but it needs a revision that fixes the split description, re-reports a single held-out test number, and addresses the count and recipe issues.

For a reading group, it's a reasonable example of applying CV to a niche public health problem, though not a methodological standout. I'd cite it only if I worked on tick classification or related applied transfer learning.

Recommendation: send to peer review, but be clear it needs heavy revision before acceptance.

Referee Report

4 major / 7 minor

Summary. The paper presents a binary image classifier that distinguishes blacklegged ticks (Ixodes scapularis) from other tick species, using 6,294 expert-identified ticks photographed at Public Health Ontario. The authors compare training from scratch, ImageNet transfer learning, attention transfer (AT), and AT combined with label smoothing regularization (LSR), reporting a best test accuracy of 92.55% for AT+LSR with a compact 5.3-million-parameter student CNN. They also describe a web application for external validation and for integrating tick identification with geography-based Lyme disease risk assessment.

Significance. If the reported accuracy reflects generalization to previously unseen individual ticks, the paper demonstrates a practical, deployable deep-learning tool for tick species identification, with a compact model size that is suitable for mobile or web deployment. The dataset of 6,294 expert-labeled ticks with dorsal and ventral images and the teacher-student knowledge-transfer scheme are potentially valuable resources for the infectious-disease diagnostic community. However, the manuscript currently lacks a clear statement of the train/test split unit, contains inconsistent dataset counts, and does not fully specify how the reported test accuracy was obtained relative to the k-fold cross-validation protocol. These issues are load-bearing because the central claim (92% accuracy, useful for prophylaxis decisions) rests on the model's ability to generalize to new ticks, not to recognize individual ticks seen during training. The web application and the explicit goal of external validation are strengths, but the evaluation protocol must be clarified before the accuracy claim can be accepted.

major comments (4)

[Section IV (and Section III-A)] The manuscript never states the unit of the train/test split. Section III-A says each tick was photographed twice (dorsal and ventral) and that the dataset includes 6,294 distinct ticks; Section IV reports an '11/1' split into 12,554 training and 1,034 test images but does not say whether the split is by image or by tick. If the split is at the image level, the two photographs of the same tick can appear on opposite sides of the split, allowing the network to memorize individual ticks and inflating the reported test accuracy relative to the stated goal of identifying unseen ticks. Please specify the split unit and, if the current split is image-level, re-run the evaluation at the tick level (e.g., assign each tick's two images to the same partition) or explain why image-level splitting is sufficient for the paper's generalization claim.
[Section IV (dataset counts)] The dataset counts are internally inconsistent. Section III-A states that 12,588 images were captured (two per tick for 6,294 ticks) and that an additional 1,000 high-resolution images were taken, which sums to 13,588 images. The split sizes reported in Section IV (12,554 training + 1,034 test) also sum to 13,588, but the text describes the total as 12,588 and calls the split an '11/1' ratio; an 11/1 split of 13,588 images would give approximately 1,132 test images, not 1,034. Please reconcile these numbers and state the exact composition of the train and test sets, including how the 1,000 high-resolution images are distributed.
[Section IV (evaluation protocol)] The evaluation protocol is unclear. Section IV says 'K-fold (k=3) cross validation was used for model evaluation and hyper-parameter tuning on the validation set,' but Tables I and II report accuracies with standard deviations on what is called a 'test set,' and Figure 4 is described as the confusion matrix 'on the test set.' It is not clear whether the reported numbers are averages over the three validation folds or come from a separate held-out test set, nor whether the same folds used for hyper-parameter selection were also used for the final accuracy report. Please clarify the exact roles of the 11/1 split and the k-fold cross-validation, and report which split the numbers in Tables I, II, and Figure 4 correspond to.
[Section V (comparison claim)] The statement in Section V that the AT+LSR model 'outperforms other models in terms of accuracy' is not supported by the reported numbers. Table I lists Lighter CNN random at 91.68 ± 0.25 and Inception-Resnet random at 92.04 ± 0.48; Table II lists AT+LSR at 92.55 ± 0.39. The differences between AT+LSR and the top Table I models are within the reported standard deviations, and AT alone (91.20 ± 0.33) is numerically lower than the lighter CNN trained from scratch (91.68 ± 0.25). Please either temper this claim or provide a formal statistical comparison (e.g., paired tests or confidence intervals on the accuracy differences).

minor comments (7)

[Table II heading] The heading contains a typo: 'specious' should be 'species.'
[Section II, paragraph 3] In the sentence about Zagoruyko et al., 'it's output' should be 'its output.'
[Section V, paragraph 2] The text says the network has '5.3 trainable parameters' but should read '5.3 million trainable parameters.'
[Section III-B, Eq. (3)] Equation (3) is ambiguous: the first term appears to be a weighted cross-entropy, but the notation '-1/β1 ∑(pi log qi)' is unclear because the minus sign and the weighting are combined. Please write the cross-entropy term explicitly and define the range of the summation.
[Section III-B, LSR description] The replacement of predicted probabilities for incorrectly classified images with a constant probability of 0.6 is a non-standard modification of label smoothing. Please justify this choice and report sensitivity to the value (e.g., a small ablation).
[Figure 2 caption] The caption says the teacher's and student's channel dimensions are 1536 and 32, respectively, but the student network architecture in Table III contains layers with 64, 128, 256, and 32 channels. Please specify which layer corresponds to the attention transfer point and clarify the channel counts.
[Abstract and Section V] The abstract and discussion emphasize integration with the geography of exposure for Lyme disease risk assessment, but the paper does not describe how the CNN output is combined with geographic data. If this integration is future work, state so explicitly in the body of the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the 92% accuracy is an empirical measurement on a held-out test split, and AT/LSR are standard methods with no parameter fitted to the reported accuracy.

full rationale

The paper's central claim is an empirical classification result, not a derived prediction that reduces to its own inputs. Section III-B defines the attention-transfer loss (Eq. 2) and the total AT+LSR loss (Eq. 3) using standard formulas from Zagoruyko et al. and Szegedy et al.; no target quantity (e.g., test accuracy) appears as a defining input. Section IV reports the AT+LSR accuracy as 92.55 ± 0.39 on a 1,034-image test set, which is a measurement rather than a fitted constant renamed as a prediction. The teacher is an ImageNet pre-trained Inception-ResNet, not a model fitted to the tick test labels, and no model parameter is optimized against the reported test accuracy. Reference [30] is a same-author preprint cited alongside [29], but no load-bearing claim in the paper depends on a result unique to [30]; the AT procedure is explicitly attributed to Zagoruyko et al. The most plausible concern is evaluation validity, not circularity: Section III-A says each tick has dorsal and ventral photos, while Section IV reports an 11/1 split 'without any overlap' without stating whether the split unit is an image or a tick. If the split is image-level, the same tick could appear in both training and test, inflating accuracy through individual-level memorization. That is a potential leakage/correctness issue, not an instance of self-definition, fitted-input-called-prediction, or self-citation load-bearing circularity under the given rubric. The paper also self-identifies the limitation that the white image background may bias the CNN, which is a generalization concern rather than a circular step. Therefore no circular step can be exhibited by construction, and the appropriate finding is no significant circularity.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on expert labels, controlled imaging conditions, and hand-set hyperparameters; no new physical or biological entities are introduced.

free parameters (5)

beta_1 (LSR cross-entropy weight) = 1
Equation (3) weights the soft-label cross-entropy term; value is hand-set in Section IV.
beta_2 (attention loss weight) = 2
Equation (3) weights the attention-transfer loss; value is hand-set in Section IV.
temperature T in soft-label generation = 5
Higher T produces softer probability distributions for student 1's outputs; set in Section IV.
replacement probability for incorrectly classified images = 0.6
Student 1's soft labels for misclassified images are replaced with a constant class probability of 0.6 (Section III-B).
image augmentation zoom range = 0.5x to 2x
Random zoom augmentation range chosen for training data in Section IV; affects model robustness.

assumptions (4)

domain assumption Expert-determined labels from Public Health Ontario are reliable ground truth for tick species.
The models are trained and evaluated against manual expert identifications under a stereomicroscope (Section III-A); if these labels are wrong, the reported accuracy is uninterpretable.
domain assumption The controlled imaging setup (phone mounted 8 cm above ticks on white paper) captures enough discriminating morphology for automatic classification, and this distribution matches deployment conditions.
Section III-A describes image acquisition; Section V admits the white background may bias the model. The intended phone app assumes field photos will be similar.
domain assumption The tick population in the dataset (Ontario, May to November 2019, 41% blacklegged) is representative of the population where the tool will be used.
Section III-A lists only one season and one province; geographic and seasonal variation in tick appearance could change accuracy.
standard math Standard deep learning optimization assumptions, e.g., Adam converges and held-out accuracy estimates generalization, hold.
Section IV uses Adam with stated hyperparameters; no proofs are needed but this is a background assumption for the empirical method.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Computer Vision Approach to Combat Lyme Disease." pith.science (2026). https://pith.science/paper/DVYJDLG4

@misc{pith2026200911931,
  author       = {Pith},
  title        = {Pith review of: A Computer Vision Approach to Combat Lyme Disease},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DVYJDLG4}},
  note         = {Machine review of arXiv:2009.11931}
}

read the original abstract

Lyme disease is an infectious disease transmitted to humans by a bite from an infected Ixodes species (blacklegged ticks). It is one of the fastest growing vector-borne illness in North America and is expanding its geographic footprint. Lyme disease treatment is time-sensitive, and can be cured by administering an antibiotic (prophylaxis) to the patient within 72 hours after a tick bite by the Ixodes species. However, the laboratory-based identification of each tick that might carry the bacteria is time-consuming and labour intensive and cannot meet the maximum turn-around-time of 72 hours for an effective treatment. Early identification of blacklegged ticks using computer vision technologies is a potential solution in promptly identifying a tick and administering prophylaxis within a crucial window period. In this work, we build an automated detection tool that can differentiate blacklegged ticks from other ticks species using advanced deep learning and computer vision approaches. We demonstrate the classification of tick species using Convolution Neural Network (CNN) models, trained end-to-end from tick images directly. Advanced knowledge transfer techniques within teacher-student learning frameworks are adopted to improve the performance of classification of tick species. Our best CNN model achieves 92% accuracy on test set. The tool can be integrated with the geography of exposure to determine the risk of Lyme disease infection and need for prophylaxis treatment.

Figures

Figures reproduced from arXiv: 2009.11931 by the authors.

**Figure 1.** (a) High-resolution microscopic images, (b) Mobile phone images of fully engorged ticks, and (c) Mobile phone images of unfed ticks. Fully and slightly engorged ticks can triple in volume when filled with blood. III. METHODS In the following section, we describe the CNN classifiers we built to detect blacklegged ticks versus other tick species. We also present the web application we developed to deploy the CNN model… view at source ↗

**Figure 2.** An overview of the Attention Transfer (AT) loss in a teacher-student learning setting. The spatial attention map is constructed by taking the sum of absolute values of a layer’s 3D tensor, A, across the channel dimension. In this setting, knowledge is transferred from the last layer of the teacher network to the one before the last layer of the student network. In the shown example, the spatial attention map Q ∈ R H… view at source ↗

**Figure 3.** The system architecture of deploying our CNN model on the web application for early identification of blacklegged ticks. For the frontend, HTML (HyperText Markup Language) and CSS (Cascading Style Sheets) were used to create the user interface (UI). On the backend, Python Flask application was developed to handle the get and post requests between UI and compute engine. Our trained CNN model was deployed on the app e… view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Confusion matrix for the best performing model which is the combination of attention transfer and label smoothing regularization on the test set. accuracy and generalization of our CNN model. Our best classification model was able to identify blacklegged ticks with 92%…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 31 canonical work pages

[1]

G. P. Wormser, R. J. Dattwyler, E. D. Shapiro, J. J. Halperin, A. C. Steere, M. S. Klempner, P. J. Krause, J. S. Bakken, F. Strle, G. Stanek, L. Bockenstedt, D. Fish, J. S. Dumler, and R. B. Nadelman, “The clinical assessment, treatment, and prevention of lyme disease, human granulocytic anaplasmosis, and babesiosis: Clinical practice guidelines by the in...

work page 2006
[2]

Lyme disease and the orthopaedic implications of lyme arthritis,

B. G. Smith, A. I. Cruz Jr, M. D. Milewski, and E. D. Shapiro, “Lyme disease and the orthopaedic implications of lyme arthritis,” The Journal of the American Academy of Orthopaedic Surgeons , vol. 19, no. 2, p. 91, 2011

work page 2011
[3]

Lyme borreliosis: a review of data on transmission time after tick attachment,

M. J. Cook, “Lyme borreliosis: a review of data on transmission time after tick attachment,” International journal of general medicine , vol. 8, p. 1, 2015

work page 2015
[4]

Ticks and tickborne bacterial diseases in humans: an emerging infectious threat,

P. Parola and D. Raoult, “Ticks and tickborne bacterial diseases in humans: an emerging infectious threat,” Clinical infectious diseases , vol. 32, no. 6, pp. 897–928, 2001. 6

work page 2001
[5]

Matrix-assisted laser desorption ionization–time of ﬂight mass spectrometry for rapid identiﬁcation of tick vectors,

A. Yssouf, C. Flaudrops, R. Drali, T. Kernif, C. Socolovschi, J.-M. Berenger, D. Raoult, and P. Parola, “Matrix-assisted laser desorption ionization–time of ﬂight mass spectrometry for rapid identiﬁcation of tick vectors,” Journal of clinical microbiology , vol. 51, no. 2, pp. 522–528, 2013

work page 2013
[6]

Increased risk of tick-borne diseases with climate and environmental changes,

C. Bouchard, A. Dibernardo, J. Kofﬁ, H. Wood, P. Leighton, and L. Lindsay, “Increased risk of tick-borne diseases with climate and environmental changes,” 2019. [Online]. Available: https://doi.org/10.14745/ccdr.v45i04a02

work page doi:10.14745/ccdr.v45i04a02 2019
[7]

CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison

J. Irvin, P. Rajpurkar, M. Ko, Y . Yu, S. Ciurea-Ilcus, C. Chute, H. Marklund, B. Haghgoo, R. Ball, K. Shpanskaya, J. Seekins, D. A. Mong, S. S. Halabi, J. K. Sandberg, R. Jones, D. B. Larson, C. P. Langlotz, B. N. Patel, M. P. Lungren, and A. Y . Ng, “CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison,” arXiv:1901.07...

work page Pith review arXiv 1901
[8]

CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning

P. Rajpurkar, J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. Langlotz, K. Shpanskaya, M. P. Lungren, and A. Y . Ng, “CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning,” arXiv:1711.05225 [cs, stat] , Nov. 2017, arXiv: 1711.05225. [Online]. Available: http://arxiv.org/abs/1711.05225

work page Pith review arXiv 2017

Show all 34 references

[9]

Deep learning for chest radiograph diagnosis: A retrospective comparison of the CheXNeXt algorithm to practicing radiologists,

P. Rajpurkar, J. Irvin, R. L. Ball, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. P. Langlotz, B. N. Patel, K. W. Yeom, K. Shpanskaya, F. G. Blankenberg, J. Seekins, T. J. Amrhein, D. A. Mong, S. S. Halabi, E. J. Zucker, A. Y . Ng, and M. P. Lungren, “Deep learning...

2018 doi
[10]

ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classiﬁcation and Localization of Common Thorax Diseases,

X. Wang, Y . Peng, L. Lu, Z. Lu, M. Bagheri, and R. M. Summers, “ChestX-ray8: Hospital-Scale Chest X-Ray Database and Benchmarks on Weakly-Supervised Classiﬁcation and Localization of Common Thorax Diseases,” in Computer Vision and Pattern Recognition (CVPR) 2017 . IEEE, 2017,...

2017
[11]

Chex- clusion: Fairness gaps in deep chest x-ray classiﬁers,

L. Seyyed-Kalantari, G. Liu, M. McDermott, and G. Marzyeh, “Chex- clusion: Fairness gaps in deep chest x-ray classiﬁers,” arXiv preprint arXiv:2003.00827, 2020

2003 arXiv
[12]

Learning to diagnose from scratch by exploiting dependencies among labels,

L. Yao, E. Poblenz, D. Dagunts, B. Covington, D. Bernard, and K. Lyman, “Learning to diagnose from scratch by exploiting dependencies among labels,” arXiv:1710.10501 [cs], Oct. 2017, arXiv: 1710.10501. [Online]. Available: http://arxiv.org/abs/1710.10501

2017 arXiv
[13]

Dermatologist-level classiﬁcation of skin cancer with deep neural networks,

A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun, “Dermatologist-level classiﬁcation of skin cancer with deep neural networks,” Nature, vol. 542, no. 7639, pp. 115–118, Feb
[14]

Clinically applicable deep learning for diagnosis and referral in retinal disease,

J. De Fauw, J. R. Ledsam, B. Romera-Paredes, S. Nikolov, N. Tomasev, S. Blackwel, H. Askham, X. Glorot, B. O’Donoghue, and D. e. a. Visentin, “Clinically applicable deep learning for diagnosis and referral in retinal disease,” Nature medicine, vol. 24, p. 1342, 2018

2018
[15]

A deep learning model to predict a diagnosis of alzheimer disease by using 18f-fdg pet of the brain,

Y . Ding, J. H. Sohn, M. G. Kawczynski, H. Trivedi, R. Harnish, N. Jenkins, D. Lituiev, T. Copeland, M. S. Aboian, C. M. Aparici, S. C. Behr, R. R. Flavell, S.-Y . Huang, K. A. Zalocusky, L. Nardo, Y . Seo, R. A. Hawkins, M. Hernandez Pampaloni, D. Hadley, and B. L. Franc, “A ...

2018
[16]

Robust automated assessment of human blastocyst quality using deep learning

P. Khosravi, E. Kazemi, Q. Zhan, M. Toschi, J. E. Malmsten, C. Hickman, M. Meseguer, Z. Rosenwaks, O. Elemento, N. Zaninovic, and I. Haji- rasouliha, “Robust automated assessment of human blastocyst quality using deep learning.”
[17]

Computer vision for microscopy diagnosis of malaria,

F. B. Tek, A. G. Dempster, and I. Kale, “Computer vision for microscopy diagnosis of malaria,” Malaria journal, vol. 8, no. 1, p. 153, 2009

2009
[18]

Using deep learning for image-based plant disease detection,

S. P. Mohanty, D. P. Hughes, and M. Salathé, “Using deep learning for image-based plant disease detection,” Frontiers in plant science , vol. 7, p. 1419, 2016

2016
[19]

Distinguishing obstructive versus central apneas in infrared video of sleep using deep learning: Validation study,

S. Akbarian, N. Montazeri Ghahjaverestan, A. Yadollahi, and B. Taati, “Distinguishing obstructive versus central apneas in infrared video of sleep using deep learning: Validation study,” J Med Internet Res, vol. 22, no. 5, p. e17252, May 2020. [Online]. Available: http://www.j...

2020
[20]

Learning what and where to transfer,

Y . Jang, H. Lee, S. J. Hwang, and J. Shin, “Learning what and where to transfer,” in ICML, 2019

2019
[21]

ImageNet: A Large-Scale Hierarchical Image Database,

J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei, “ImageNet: A Large-Scale Hierarchical Image Database,” in CVPR09, 2009

2009
[22]

A survey on transfer learning,

S. J. Pan and Q. Yang, “A survey on transfer learning,”IEEE Transactions on Knowledge and Data Engineering , vol. 22, no. 10, pp. 1345–1359, 2010

2010
[23]

Using pre-training can improve model robustness and uncertainty,

D. Hendrycks, K. Lee, and M. Mazeika, “Using pre-training can improve model robustness and uncertainty,” 2019

2019
[24]

Assessment of the generalization of learned image reconstruction and the potential for transfer learning,

F. Knoll, K. Hammernik, E. Kobler, T. Pock, M. P. Recht, and D. K. Sodickson, “Assessment of the generalization of learned image reconstruction and the potential for transfer learning,” Magnetic Resonance in Medicine , vol. 81, no. 1, pp. 116–128, 2019. [Online]. Available: ht...

2019 doi
[25]

Transfusion: Understanding Transfer Learning for Medical Imaging,

M. Raghu, C. Zhang, J. Kleinberg, and S. Bengio, “Transfusion: Understanding Transfer Learning for Medical Imaging,” in Proceedings of the 33rd International Conference on Neural Information Processing Systems, 2019

2019
[26]

Model compression,

C. Bucilu ˇa, R. Caruana, and A. Niculescu-Mizil, “Model compression,” in Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining , 2006, pp. 535–541

2006
[27]

Distilling the knowledge in a neural network,

G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” arXiv preprint arXiv:1503.02531 , 2015

2015 arXiv
[28]

Rethinking the inception architecture for computer vision,

C. Szegedy, V . Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna, “Rethinking the inception architecture for computer vision,” pp. 2818–2826, 2015

2015
[29]

Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer,

S. Zagoruyko and N. Komodakis, “Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer,” CoRR, vol. abs/1612.03928, 2016. [Online]. Available: http://arxiv.org/abs/1612.03928

2016 arXiv
[30]

Evaluating knowledge transfer in neural network for medical images,

S. Akbarian, L. Seyyed-Kalantari, F. Khalvati, and E. Dolatabadi, “Evaluating knowledge transfer in neural network for medical images,” arXiv preprint arXiv:2008.13574 , 2020

2008 arXiv
[31]

Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks,

L. Wang and K.-J. Yoon, “Knowledge distillation and student-teacher learning for visual intelligence: A review and new outlooks,” arXiv preprint arXiv:2004.05937, 2020

2004 arXiv
[32]

Inception-v4, inception-resnet and the impact of residual connections on learning,

C. Szegedy, S. Ioffe, V . Vanhoucke, and A. Alemi, “Inception-v4, inception-resnet and the impact of residual connections on learning,” 2016

2016
[33]

Distilling the knowledge in a neural network,

G. E. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” 2015. 7 APPENDIX A In the appendix, we present supplementary details about lighter CNN model as listed in Table III. The network has 13 layers with 5,350,633 trainable parameters composed of...

2015
[2017]

Available: https://www.nature.com/articles/nature21056

[Online]. Available: https://www.nature.com/articles/nature21056

Pith tools

Reviewed August 27, 2026 · model on record in the stance chip above.

[1] [1]

G. P. Wormser, R. J. Dattwyler, E. D. Shapiro, J. J. Halperin, A. C. Steere, M. S. Klempner, P. J. Krause, J. S. Bakken, F. Strle, G. Stanek, L. Bockenstedt, D. Fish, J. S. Dumler, and R. B. Nadelman, “The clinical assessment, treatment, and prevention of lyme disease, human granulocytic anaplasmosis, and babesiosis: Clinical practice guidelines by the in...

work page 2006

[2] [2]

Lyme disease and the orthopaedic implications of lyme arthritis,

B. G. Smith, A. I. Cruz Jr, M. D. Milewski, and E. D. Shapiro, “Lyme disease and the orthopaedic implications of lyme arthritis,” The Journal of the American Academy of Orthopaedic Surgeons , vol. 19, no. 2, p. 91, 2011

work page 2011

[3] [3]

Lyme borreliosis: a review of data on transmission time after tick attachment,

M. J. Cook, “Lyme borreliosis: a review of data on transmission time after tick attachment,” International journal of general medicine , vol. 8, p. 1, 2015

work page 2015

[4] [4]

Ticks and tickborne bacterial diseases in humans: an emerging infectious threat,

P. Parola and D. Raoult, “Ticks and tickborne bacterial diseases in humans: an emerging infectious threat,” Clinical infectious diseases , vol. 32, no. 6, pp. 897–928, 2001. 6

work page 2001

[5] [5]

Matrix-assisted laser desorption ionization–time of ﬂight mass spectrometry for rapid identiﬁcation of tick vectors,

A. Yssouf, C. Flaudrops, R. Drali, T. Kernif, C. Socolovschi, J.-M. Berenger, D. Raoult, and P. Parola, “Matrix-assisted laser desorption ionization–time of ﬂight mass spectrometry for rapid identiﬁcation of tick vectors,” Journal of clinical microbiology , vol. 51, no. 2, pp. 522–528, 2013

work page 2013

[6] [6]

Increased risk of tick-borne diseases with climate and environmental changes,

C. Bouchard, A. Dibernardo, J. Kofﬁ, H. Wood, P. Leighton, and L. Lindsay, “Increased risk of tick-borne diseases with climate and environmental changes,” 2019. [Online]. Available: https://doi.org/10.14745/ccdr.v45i04a02

work page doi:10.14745/ccdr.v45i04a02 2019

[7] [7]

CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison

J. Irvin, P. Rajpurkar, M. Ko, Y . Yu, S. Ciurea-Ilcus, C. Chute, H. Marklund, B. Haghgoo, R. Ball, K. Shpanskaya, J. Seekins, D. A. Mong, S. S. Halabi, J. K. Sandberg, R. Jones, D. B. Larson, C. P. Langlotz, B. N. Patel, M. P. Lungren, and A. Y . Ng, “CheXpert: A Large Chest Radiograph Dataset with Uncertainty Labels and Expert Comparison,” arXiv:1901.07...

work page Pith review arXiv 1901

[8] [8]

CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning

P. Rajpurkar, J. Irvin, K. Zhu, B. Yang, H. Mehta, T. Duan, D. Ding, A. Bagul, C. Langlotz, K. Shpanskaya, M. P. Lungren, and A. Y . Ng, “CheXNet: Radiologist-Level Pneumonia Detection on Chest X-Rays with Deep Learning,” arXiv:1711.05225 [cs, stat] , Nov. 2017, arXiv: 1711.05225. [Online]. Available: http://arxiv.org/abs/1711.05225

work page Pith review arXiv 2017