CANDOR Offers a Head-Free Test for Medical Imaging Encoders

A paper on arXiv proposes a way to compare frozen image encoders for medical-imaging tasks before training a task-specific classifier. Its measure, called chance-calibrated neighborhood discordance (CANDOR), checks whether positive images sit closer to other positive examples than to negative ones in an encoder’s feature space. The authors report results across 22 encoders, 605,443 images and eight tasks in seven domains.

The practical question is whether this measure can help researchers and hospitals shortlist encoders more cheaply. The paper says calculating its discordance rate requires a forward pass for each image, rather than first fitting a classification head. That makes it a proposed screening measure, not evidence that a particular encoder is ready for clinical use.

Key facts

What CANDOR measures

The method starts with image features produced by a frozen encoder. For each positive image, it compares the average similarity to its k nearest positive examples with the average similarity to its k nearest negative examples. A positive is counted as discordant when the negative neighbors are more similar.

The resulting rate, D, is the share of positive images that meet that condition. Under the paper’s stated setup, when the positive and negative candidate sets are equal in size, features with no information about the label have a chance-level rate of one half. A rate below one half indicates that positive images more often have closer positive neighbors than negative ones.

This differs from evaluating a trained head with AUROC. The authors argue that AUROC measures a classifier’s performance on encoder features, while CANDOR directly probes neighborhood separation in those features. The paper does not establish that the new measure should replace AUROC or other evaluation methods.

Why sampling changes the result

The study’s chest-imaging results illustrate a key caveat: candidate-set construction affects the reported rate. Across 12 findings in MIMIC-CXR and CheXpert, the authors report that all 11 chest encoders had rates above one half when the candidate sets were unequal, ranging from 0.732 to 0.927. With equal-sized sets, the same encoders had rates below one half, from 0.277 to 0.463.

The authors also report that rates increased as findings became rarer. Readers comparing encoder scores therefore need to know how examples were sampled and whether the positive and negative sets were balanced. A number without that context could give a misleading impression of model quality.

The paper says linear heads missed 0.301 of discordant positive examples on the chest test split, compared with 0.165 of the other positives. This is a reported association in the study, not proof that CANDOR alone identifies clinically important errors or that a hospital adopting it would improve patient outcomes.

A screening step, not a clinical verdict

For AI researchers and medical-imaging teams, the proposed benefit is timing: use a feature-level measure to narrow the encoder candidates before investing effort in training task-specific heads. The authors say the calculation takes one forward pass per image. The supplied paper summary does not provide enough detail to assess hardware costs, runtime in a hospital workflow, or how the method behaves under changes in scanners, patient populations and labeling practice.

Those limits matter because a favorable neighborhood score does not by itself establish diagnostic accuracy, safety, or suitability for deployment. The reported results span several tasks and domains, but the available source material does not specify enough about each dataset, comparison protocol or uncertainty estimate to independently judge the strength of every result.

What to check next

Teams considering CANDOR should inspect the paper’s full methods and results, particularly the definition of candidate sets, choice of k, handling of class imbalance and evaluation splits. They should also compare its rankings with task-specific validation results on data representative of their intended setting.

The arXiv record identifies the paper as version 4 and gives a publication date of October 27, 2026. The available material does not include independent replication or external commentary, so the reported findings should be treated as the authors’ results pending further validation.

Source: arXiv cs.CL, “CANDOR: Chance-Calibrated Neighborhood Discordance in Frozen Encoders for Medical Imaging” — https://arxiv.org/abs/2607.18451