OrganAMNIST: Abdominal Imaging Benchmark

• OrganAMNIST is a curated medical imaging benchmark from MedMNIST featuring small grayscale abdominal CT/MRI images and an 11-class organ taxonomy.
• It supports diverse experiments, including studies on architectural inductive bias, transfer learning, federated learning, and adversarial robustness using standardized splits.
• The dataset enables evaluations of uncertainty, privacy, and inverse imaging methods, providing actionable insights for robust medical imaging model development.

OrganAMNIST is a MedMNIST/MedMNIST v2 medical-imaging benchmark that recent literature uses primarily for abdominal organ classification, while also treating it as a compact testbed for architectural inductive bias, representation transfer, federated learning, machine unlearning, conformal prediction, adversarial robustness, domain adaptation, quantum-classical learning, and optical image reconstruction. Most of the cited work describes OrganAMNIST as grayscale abdominal CT slices, whereas one machine-unlearning study describes it as axial abdominal MRI, so modality and related metadata are best interpreted paper-by-paper rather than as perfectly uniform across all uses (Shahjalal et al., 16 Sep 2025, Falcao et al., 25 Aug 2025).

1. Dataset identity and reported specifications

Across the cited papers, OrganAMNIST is consistently framed as an 11-class single-label classification problem over small grayscale medical images drawn from MedMNIST or MedMNIST v2. Several papers report the same nominal dataset size and split—$58{,}830$ total samples with $34{,}561$ train, $6{,}491$ validation, and $17{,}778$ test—which is the configuration used in both hybrid quantum-classical classification and machine-unlearning studies (Shahjalal et al., 16 Sep 2025, Falcao et al., 25 Aug 2025). A conformal-prediction study instead reports $58{,}850$ images with $34{,}581$ train, $6{,}491$ validation, and $17{,}778$ test, making explicit that paper-specific reporting is not fully identical (Octadion et al., 13 May 2026).

The label set is also stable in substance but not always in wording. One paper lists the 11 classes as Bladder, Femur-Left, Femur-Right, Heart, Kidney-Left, Kidney-Right, Liver, Lung-Left, Lung-Right, Pancreas, and Spleen (Octadion et al., 13 May 2026). Another lists right femoral head, heart, left femoral head, bladder, left kidney, right kidney, liver, left lung, right lung, spleen, and pancreas (Maqbool et al., 25 Nov 2025). This suggests a consistent organ taxonomy with minor naming differences around the femoral classes.

Reported modality and resolution are more variable. OrganAMNIST is described as abdominal CT in multiple studies, including hybrid quantum-classical classification, structured ViT transfer, conformal prediction, artifact robustness, and multimode-fiber reconstruction (Shahjalal et al., 16 Sep 2025, Wang et al., 10 Mar 2026, Cheung et al., 8 Oct 2025, Maqbool et al., 25 Nov 2025). By contrast, the unlearning study describes it as containing axial slices from abdominal MRI scans (Falcao et al., 25 Aug 2025). Similarly, some papers use the MedMNIST-standard $28\times 28$ grayscale images directly (Shahjalal et al., 16 Sep 2025), some resize all inputs to $64\times 64$ (Falcao et al., 25 Aug 2025), and others upsample or convert to model-specific formats such as $34{,}561$0 RGB for ViT or ResNet pipelines (Sun et al., 2023, Octadion et al., 13 May 2026).

A particularly influential characterization comes from the compact-ViT literature, which places OrganAMNIST at the strongest end of a “spatial structure spectrum,” describing it as “stable abdominal anatomy” with spatial strength index 5 (Angelakis, 20 Feb 2026). In that usage, OrganAMNIST is not merely an organ-label benchmark; it is an example of a regime in which spatial layout is consistently informative.

Aspect	Reported values	Example source
Modality	Abdominal CT; axial abdominal MRI	(Shahjalal et al., 16 Sep 2025, Falcao et al., 25 Aug 2025)
Class count	11 classes	(Octadion et al., 13 May 2026)
Common split	34,561 / 6,491 / 17,778	(Shahjalal et al., 16 Sep 2025)
Alternate split report	34,581 / 6,491 / 17,778	(Octadion et al., 13 May 2026)
Native-style input	$34{,}561$1 grayscale	(Shahjalal et al., 16 Sep 2025)
Model-specific resizing	$34{,}561$2, $34{,}561$3, RGB conversion	(Falcao et al., 25 Aug 2025, Sun et al., 2023)

2. Experimental roles across recent literature

OrganAMNIST has become a multi-role benchmark rather than a single-task dataset. Recent work uses it to probe whether spatial priors help or hurt, whether chest-X-ray pretraining transfers to abdominal CT, whether unlabeled artifact domains can be adapted to, whether uncertainty sets remain valid under attack, and whether quantum or hybrid optical systems can remain useful on medically structured data.

Study	Setting	OrganAMNIST role
ZACH-ViT	Few-shot ViTs	Strong anatomical-prior regime
VIVID-Med	Cross-modality transfer	CT organ benchmark
SalUn	Machine unlearning	Medical privacy benchmark
FedPerfix	Personalized federated ViTs	Medical FL benchmark
Adaptive RAPS	Reliable set prediction	Primary radiology benchmark
HistoSpeckle-Net	MMF reconstruction	Structurally rich target images

This breadth matters because different papers expose different properties of the dataset. In architecture papers, OrganAMNIST is usually a high-structure multi-class problem. In privacy and federated studies, it is a medically relevant dataset with enough size to support retraining, unlearning, and client partitioning. In uncertainty and robustness work, it functions as a compact but nontrivial abdominal-imaging domain where failure modes remain measurable. In inverse-imaging work, it becomes a source of anatomically rich targets whose internal textures and boundaries are difficult to reconstruct from degraded measurements (Angelakis, 20 Feb 2026, Wang et al., 10 Mar 2026, Falcao et al., 25 Aug 2025, Sun et al., 2023, Octadion et al., 13 May 2026, Maqbool et al., 25 Nov 2025).

3. Spatial structure, transfer, and architectural inductive bias

A central result in the recent OrganAMNIST literature is that the dataset rewards spatially aware inductive bias more than permutation-invariant designs. In the ZACH-ViT study, OrganAMNIST is explicitly assigned to the “strong anatomical prior” regime. Under a strict few-shot protocol with 50 samples per class, fixed hyperparameters, batch size 16, learning rate $34{,}561$4, AdamW, 23 epochs, and seeds $34{,}561$5, ZACH-ViT reaches $34{,}561$6 MacroF1 on OrganAMNIST, compared with $34{,}561$7 for Minimal-ViT and $34{,}561$8 for TransMIL (Angelakis, 20 Feb 2026). The paper interprets this as evidence that removing positional embeddings and the $34{,}561$9 token is less advantageous when “spatial relationships are diagnostically relevant.”

That interpretation aligns with OrganAMNIST’s repeated use as a transfer benchmark for structured visual representations. In VIVID-Med, a vit_base_patch16_224 encoder pretrained only on chest X-rays with LLM-supervised structured semantics is frozen and evaluated by linear probing for 3,000 steps. On OrganAMNIST, this yields $6{,}491$0 macro-AUC and $6{,}491$1 macro-F1, exceeding ImageNet-supervised ViT at $6{,}491$2 macro-AUC and $6{,}491$3 macro-F1, and BiomedCLIP at $6{,}491$4 macro-AUC and $6{,}491$5 macro-F1 (Wang et al., 10 Mar 2026). Here OrganAMNIST functions as evidence that structured supervision learned on chest radiographs can transfer to abdominal CT without CT pretraining.

Hybrid quantum-classical classification studies make a related point from a different direction. HQCNN, using a five-layer classical convolutional backbone plus a 4-qubit variational quantum circuit with angle embedding, cyclic $6{,}491$6, superpositional entanglement, and a Quantum Attention-Fourier layer, reports $6{,}491$7 accuracy and $6{,}491$8 AUC on the 11-class OrganAMNIST task (Shahjalal et al., 16 Sep 2025). Its ablation study shows that removing QAF reduces accuracy from $6{,}491$9 to $17{,}778$0 and AUC from $17{,}778$1 to $17{,}778$2, indicating that the quantum module’s most elaborate component contributes measurably on organ discrimination.

A separate hybrid QCNN paper uses only four OrganAMNIST classes, indexed $17{,}778$3, and shows that reusing discarded qubit measurements can markedly improve performance on this harder, class-imbalanced biomedical subset. In its best reported configuration, accuracy rises from $17{,}778$4 without discarded-qubit reuse to $17{,}778$5 with reuse, while F1 rises from $17{,}778$6 to $17{,}778$7 (Anwar et al., 25 Aug 2025). Although this is not the full 11-class setting, it reinforces the broader pattern: OrganAMNIST is repeatedly used to test whether a model captures structured anatomical information rather than only coarse appearance cues.

4. Privacy, federation, and distribution shift

OrganAMNIST has also become a compact benchmark for operational questions that go beyond standard supervised learning. In machine unlearning, a ResNet-18 trained for 200 epochs with learning rate $17{,}778$8, batch size $17{,}778$9, random crop, and horizontal flip is used to compare SalUn against full retraining under forgetting rates $58{,}850$0 and $58{,}850$1. At $58{,}850$2, retraining reaches $58{,}850$3 test accuracy and $58{,}850$4 minutes runtime, while SalUn reaches $58{,}850$5 test accuracy and $58{,}850$6 minutes runtime; at $58{,}850$7, retraining reaches $58{,}850$8 and SalUn $58{,}850$9, with runtime $34{,}581$0 versus $34{,}581$1 minutes (Falcao et al., 25 Aug 2025). OrganAMNIST is therefore one of the clearest medical-image cases in which approximate unlearning approaches full retraining while sharply reducing computation.

In personalized federated learning, OrganAMNIST serves as the medical-imaging benchmark for partial ViT personalization. FedPerfix partitions the dataset across $34{,}581$2 clients using a Dirichlet split with $34{,}581$3, resizes inputs to $34{,}581$4 RGB, uses ViT-Small with patch size 16, runs 50 communication rounds with 10 local epochs and $34{,}581$5 sampled clients per round, and reports mean client-wise Top-1 accuracy (Sun et al., 2023). On this setup, FedPerfix achieves $34{,}581$6, exceeding FedAvg at $34{,}581$7 and FedRep at $34{,}581$8. OrganAMNIST is thus used to show that self-attention and the classification head are sufficiently distribution-sensitive to justify local plugins.

Quantum federated learning uses a smaller OrganAMNIST subset but addresses a related heterogeneity problem. Q-RAIL employs 4,700 training samples and 1,000 test samples, reduces images by PCA to $34{,}581$9 features, and trains a 4-qubit, 4-layer VQC across $6{,}491$0 clients for 15 rounds using SPSA. Under IID partitioning, Q-RAIL improves test accuracy from FedAvg’s $6{,}491$1 to $6{,}491$2; under non-IID partitioning, it improves from $6{,}491$3 to $6{,}491$4 (Maouaki et al., 25 May 2026). In that paper, OrganAMNIST is the medical domain used to show that hardware-aware aggregation helps even when the task is more structured than MNIST.

Distribution shift caused by CT artifacts is another prominent use. A domain-adaptation study builds clean, uniform-noise, 90°-rotated, and ring-artifact versions of OrganAMNIST, then trains a ResNet-50-based DANN with a feature extractor $6{,}491$5, label head $6{,}491$6, domain classifier $6{,}491$7, and gradient reversal $6{,}491$8. The total loss is

$6{,}491$9

so only clean source samples contribute to supervised class loss, while both clean and artifact samples contribute to domain loss (Cheung et al., 8 Oct 2025). For ring-artifact adaptation, the best $17{,}778$0 schedule is parabolic increasing, with validation accuracy $17{,}778$1. The paper’s main qualitative result is that baseline models trained only on clean images fail on ring-artifact test data, ordinary augmentation does not help on unseen artifact domains, and DANN with unlabeled artifact images attains ring-artifact performance comparable to models trained with labeled artifact images (Cheung et al., 8 Oct 2025).

5. Reliability, conformal prediction, and adversarial robustness

OrganAMNIST is unusually prominent in recent reliability work because it exposes a tension between excellent average performance and localized failure. In adaptive conformal prediction, a ResNet-18 pretrained on ImageNet-1K and adapted to 11 classes reaches $17{,}778$2 test accuracy, yet standard RAPS tuned for minimal set size collapses to near-deterministic behavior. On OrganAMNIST, Naive and RAPS (Size) both have average set size $17{,}778$3 and coverage about $17{,}778$4, but RAPS (Size) has worst populated-stratum coverage $17{,}778$5, and RAPS (Temp) reduces that further to $17{,}778$6 (Octadion et al., 13 May 2026). The proposed Adaptive Lambda Criterion instead attains coverage $17{,}778$7, average set size $17{,}778$8, singleton rate $17{,}778$9, and worst populated-stratum coverage $28\times 28$0. It also produces 1,244 multi-label predictions, compared with only 10 under RAPS (Size), and the paper reports a Spearman correlation $28\times 28$1 with $28\times 28$2 between Grad-CAM spatial entropy and prediction-set size, indicating that larger sets correspond to more focused attention on anatomically ambiguous regions (Octadion et al., 13 May 2026).

A separate conformal-prediction paper studies OrganAMNIST under explicit adversarial attacks using a single-channel ResNet18 and APS scores. Under known PGD attack, the PGD-trained defensive model achieves coverage $28\times 28$3, size $28\times 28$4, SSCV $28\times 28$5, and accuracy $28\times 28$6, whereas the normal model retains similar coverage but expands to size $28\times 28$7 with accuracy $28\times 28$8 (Luo et al., 2024). Under unknown attacks, conservative quantile calibration keeps robust-model set sizes around $28\times 28$9–$64\times 64$0, while the normal model produces sets around $64\times 64$1–$64\times 64$2. The paper’s game-theoretic analysis further concludes that the OrganAMNIST minimax defense often degenerates to a single robust model, frequently the PGD-trained ResNet18 (Luo et al., 2024).

Hybrid quantum-classical adversarial defense uses OrganAMNIST as its medical testbed. QShield compares a modified grayscale ResNet-18 with hybrid CNN-plus-PQC variants under FGSM, PGD, APGD, VMI-FGSM, C&W $64\times 64$3, DeepFool, One-Pixel, and Square attacks. The classical CNN reaches clean ODR $64\times 64$4, while hybrid HQCNN variants remain close on clean performance and achieve lower attack success rates across attacks (Azimi et al., 13 Apr 2026). The best reported OrganAMNIST improvements include up to $64\times 64$5 absolute ASR reduction under C&W and a relative robustness improvement of $64\times 64$6 under Square attack for HQCNN-Linear. The same study reports that adversarial-example generation becomes far more expensive against the hybrid models: on OrganAMNIST, C&W requires $64\times 64$7 s/sample for the CNN but $64\times 64$8–$64\times 64$9 s/sample for HQCNN variants, while Square attack rises from $34{,}561$00 s/sample to $34{,}561$01–$34{,}561$02 s/sample (Azimi et al., 13 Apr 2026).

These papers collectively undermine two common shortcuts in reading OrganAMNIST results. First, near-saturated AUC or high global coverage does not guarantee robustness or reliable uncertainty. Second, strong clean accuracy does not imply compact or useful predictive sets under attack. OrganAMNIST repeatedly functions as the dataset on which those distinctions remain visible.

6. Beyond classification: inverse imaging, reconstruction, and benchmark significance

OrganAMNIST has also been used as a target distribution for optical inverse problems rather than only as a classification benchmark. HistoSpeckle-Net transmits OrganAMNIST images through a multimode fiber using an SLM and records output speckles under three bending configurations, creating 58,830 speckle–image pairs per configuration (Maqbool et al., 25 Nov 2025). The generator combines a U-Net-like backbone, a Three-Scale Feature Refinement Module, histogram-based mutual-information loss, multiscale SSIM loss, and a PatchGAN discriminator. On a 5,883-image test set per configuration, HistoSpeckle-Net reaches average SSIM $34{,}561$03, compared with $34{,}561$04 for U-Net and $34{,}561$05 for Pix2Pix; with only 15,000 training images, it still reaches $34{,}561$06, compared with $34{,}561$07 and $34{,}561$08, and under combined perturbed-fiber training it maintains SSIM $34{,}561$09 for all bending positions (Maqbool et al., 25 Nov 2025). In this setting, OrganAMNIST is valued not for labels alone but for “rich structures,” “fine details,” organ boundaries, and internal textures that stress reconstruction fidelity.

The cumulative literature therefore presents OrganAMNIST as more than a small MedMNIST subset. It is a compact abdominal-imaging benchmark whose stable anatomy makes spatial priors measurable, whose ambiguities make calibrated uncertainty and adversarial robustness nontrivial, whose standardized format supports privacy and federated experiments, and whose structural richness is sufficient to test optical reconstruction and hybrid quantum models (Angelakis, 20 Feb 2026, Octadion et al., 13 May 2026, Maqbool et al., 25 Nov 2025). At the same time, several papers explicitly mark the limits of current conclusions: ZACH-ViT evaluates only a 50-shot-per-class regime (Angelakis, 20 Feb 2026); adaptive conformal work notes that OrganAMNIST is low-resolution and that some uncertainty may be artifactual (Octadion et al., 13 May 2026); artifact-robustness experiments use synthetic rather than scanner-acquired ring artifacts (Cheung et al., 8 Oct 2025); and multimode-fiber reconstruction emphasizes that OrganAMNIST remains a curated 2D single-channel dataset rather than full clinical imaging complexity (Maqbool et al., 25 Nov 2025). This suggests that OrganAMNIST is most informative when treated as a controlled, structurally meaningful benchmark whose research value lies in the precision with which different methodological assumptions can be isolated and compared.