Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2608.22059

CRS-Bench: A Reference-Relative Reliability Benchmark for Medical Image Encoders

Xingtao Lin·Hangqi Ren·Caiwan Sun·You Chen
View original source ↗Read source document ↗Export metadata ↓

Abstract

Pretrained image encoders are central to medical image classification, where expert annotation is costly and task-specific cohorts are often limited. As the model space expands from general-purpose to broad-medical and specialty-specific encoders, selecting the representation becomes a substantive modeling decision. Clean-test discrimination alone is insufficient for this purpose: encoders with similar AUROC can differ in calibration, label efficiency, and stability under acquisition perturbations or distribution shift. We introduce CRS-Bench, a controlled benchmark for multi-objective medical encoder selection. CRS-Bench evaluates 15 pretrained encoder families across dermatology, ophthalmology, and radiology using ISIC 2019, APTOS 2019, and CheXpert, with CheXpert-to-MIMIC-CXR as an observed institutional shift, yielding 17,575 controlled run records and 3,515 seed-aggregated metric rows. Each encoder is characterized along four operational reliability dimensions: discrimination, calibration, label efficiency, and robustness. We summarize these dimensions using the Clinical Reliability Score (CRS), a Pareto-aware, reference-relative score combining dominance, profile balance, and worst-axis performance. AUROC and CRS are positively associated but not decision-equivalent: 21 of 105 pairwise orderings reverse, with a mean absolute rank displacement of 1.87. Paired-seed bootstrap analysis identifies PanDerm, MedSigLIP, and MedGemma as a stable leading reliability tier rather than a statistically resolved single leader. CRS-Bench provides a controlled framework for selecting medical image encoders from multi-axis reliability profiles rather than clean-test AUROC alone.

Explore related subjects

eess.IVcs.AIcs.CV

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Xingtao Lin, Hangqi Ren, Caiwan Sun, You Chen. 2026-08-22. CRS-Bench: A Reference-Relative Reliability Benchmark for Medical Image Encoders. https://arxiv.org/abs/2608.22059

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

JEVQA - Video Quality from Metadata, Bitstream, and Pixel Features with a General-Purpose Decision Model

Instrumental quality models for video quality prediction are usually trained for a fixed set of codecs or other input features, and every new input variable requires retraining. Novel, general-purpose decision models can answer questions without task-specific training, but it is unclear whether they can judge video quality. We evaluate Jev, a commercial ``System One'' model that returns probability distributions over a provided answer scale, as a zero-shot video quality model. We call the resulting method JEVQA. In a first study on 1,936 AV1, H.264, HEVC, and VP9 encodes of 22 sources, scored against VMAF as ground truth, using encoding metadata only, JEVQA reached a Pearson correlation of 0.737, on par with the standardized ITU-T P.1204.1 model (0.733). Giving the model bitstream data raised the accuracy to 0.797, and combined pixel-based and bitstream features raised it to 0.824. A pixel-only variant failed in our tests. In a second study, using H.264, HEVC, and VP9 encodes in the AVT-VQDB-UHD-1 database, the metadata-only model reached a correlation of 0.879 with MOS, close to P.1204.1 (0.898). Bitstream statistics did not help there. Our results show that trained models on the same features remain clearly ahead in both studies, but that zero-shot classifiers are promising.

eess.IV↗
arXiv2026

When is a closed-form RGB->S/P ratio adequate? A hyperspectral characterization on natural scenes for mesopic display

Mesopic and low-light display transforms require, as their driving signal, a per-pixel scotopic-to-photopic luminance ratio (S/P); the exact spectral S/P is unavailable for ordinary RGB content, so a low-cost closed form that estimates S/P from a linear-RGB triplet is used in its place. Such closed forms exist but have been characterized only on narrowband / LED sources, i.e. spectrally sparse spectra, where a relative error of ~41% has been reported for a three-channel projection. Display content, however, is natural and broadband. We ask whether the same closed form is adequate there, using per-pixel spectral S/P from hyperspectral imagery as ground truth. On a daylight radiance time-series, a six-scalar closed form (three photopic and three scotopic channel weights) reproduces spectral S/P with a median error of ~0.07 that is time-invariant once the RGB input is chromatically adapted to D65; evaluated in un-adapted sRGB the error instead carries a color-temperature tilt across illuminants (~0.19), so adaptation is the enabling step for this use case. The result generalizes to an independent fifty-scene set (pooled median 0.024; 45/50 scenes within a pre-registered 0.10 band), with the few exceedances concentrated in saturated, spectrally-peaky surfaces that approach the narrowband regime (floral close-ups in this set). The scotopic weight vector is shown to be primary-model dependent, but the value used here is corroborated by a primary-free XYZ projection, and the median error stays within the band across all principled coefficient choices. We do not claim observer-validated appearance fidelity or adequacy on narrowband sources; both are out of scope. Both outcomes follow from the same three-channel projection: it is overwhelmed by spectrally sparse inputs and adequate on spectrally smooth ones.

eess.IV↗
arXiv2026

Recurrent Convolutional Neural Networks for LiDAR-Based Attitude Initialization of Rotating Spacecraft

Accurate attitude estimation is essential for autonomous in-orbit servicing and proximity operations. This work proposes a Recurrent Convolutional Neural Network (RCNN) used in coarse attitude initialization of known, possibly tumbling spacecraft using LiDAR-derived depth images. By processing temporal sequences of 2D point-cloud projections, the RCNN effectively handles symmetries, occlusions, and degraded sensing. Simulations across various spacecraft geometries, angular velocities, and ranges show that the RCNN yields lower initialization errors and higher convergence rates than conventional CNN baseline within the adopted experimental framework, with performance varying across angular velocity conditions.

eess.IV↗