Orochi: Ultra-bright SMG and Biomedical Image Processor

• Orochi is a dual-meaning term referring both to an ultra-bright, high-redshift submillimeter galaxy and a versatile biomedical image processor.
• In astronomy, Orochi (SXDF1100.001) is an optically dark SMG detected with high flux densities, exhibiting a two-component morphology suggestive of gravitational lensing.
• In biomedical imaging, Orochi is a unified processor that streamlines registration, fusion, restoration, and super-resolution via task-aware pretraining on large-scale datasets.

Searching arXiv for the provided Orochi papers to ground the article. Orochi is a name used for two unrelated research entities in the arXiv literature summarized here. In extragalactic astronomy, Orochi denotes the submillimeter galaxy AzTEC-ASTE-SXDF1100.001, also catalogued as SXDF1100.001, an exceptionally bright source in the Subaru/XMM-Newton Deep Field that is interpreted as an optically dark, likely gravitationally lensed high-redshift starburst (Ikarashi et al., 2010). In biomedical imaging, Orochi denotes a versatile image processor for low-level biomedical image tasks, designed as a unified framework for registration, fusion, restoration, and super-resolution (Dai et al., 26 Sep 2025).

1. Disambiguation and nomenclature

The astronomical Orochi is presented as an exceptionally bright submillimeter galaxy discovered in a wide-field AzTEC/ASTE 1100 $\mu$m survey of the SXDF. The source is explicitly identified as AzTEC-ASTE-SXDF1100.001 or SXDF1100.001 and is discussed as one of the brightest SMGs known at the time, with the preferred interpretation that it is an optically dark SMG at $z\sim3.4$ behind a foreground red galaxy at $z\sim1.4$ (Ikarashi et al., 2010).

The biomedical Orochi is presented as the first application-oriented, efficient, and versatile biomedical image processor for low-level imaging tasks. Its stated purpose is to replace fragmented collections of task-specific models and plugins in environments such as ImageJ/Fiji and napari with a single pretrained, task-aware framework (Dai et al., 26 Sep 2025).

Usage	Designation	Core description
Astronomy	AzTEC-ASTE-SXDF1100.001 / SXDF1100.001	Exceptionally bright SMG in the SXDF
Biomedical imaging	Orochi	Unified processor for registration, fusion, restoration, and super-resolution

2. Discovery and measurement of the astronomical Orochi

Orochi was discovered serendipitously in an AzTEC/ASTE 1100 $\mu$m survey. In the AzTEC map it appears as a very bright point-like source with measured flux density

$S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$

The AzTEC/ASTE beam is $30^{\prime\prime}$, with map noise of $0.6$–$1.0$ mJy, so the initial detection could not by itself distinguish between a compact source, an extended source, a blend, or a lensed system. The reported source position is $\alpha=02^h18^m30.68^s$, $\delta=-05^\circ31'31.37''$ (Ikarashi et al., 2010).

Follow-up measurements were obtained with CARMA at 1300 $z\sim3.4$0m, SMA at 880 $z\sim3.4$1m, and Z-Spec on the CSO over 190–308 GHz, together with optical, near-infrared, mid-infrared, and radio archival data. CARMA, observed in D configuration with 3 GHz total continuum bandwidth, detected the source at $z\sim3.4$2 with

$z\sim3.4$3

where the second uncertainty reflects uv-model fitting of the extended structure. SMA, in compact configuration with 8 GHz bandwidth, detected the source at $z\sim3.4$4 with

$z\sim3.4$5

This places Orochi among the brightest known SMGs and makes it comparable to or brighter than lensed sources such as SMM J2135-0102.

Z-Spec detected continuum throughout the band but no statistically significant emission or absorption lines. Its best-fit continuum is

$z\sim3.4$6

with reported continuum points $z\sim3.4$7 mJy, $z\sim3.4$8 mJy, $z\sim3.4$9 mJy, $z\sim1.4$0 mJy, $z\sim1.4$1 mJy, and $z\sim1.4$2 mJy. These measurements established both the exceptional brightness of the source and the need for a structural interpretation beyond a single unresolved component.

3. Structure, counterpart identification, and redshift interpretation

A central result of the astronomical study is that Orochi is not a simple unresolved point source. The SMA and CARMA visibility amplitudes decrease with baseline length, and the data are fit with a two-component model comprising an unresolved compact component and an extended Gaussian component (Ikarashi et al., 2010).

At 880 $z\sim1.4$3m, the SMA fit gives a compact component with $z\sim1.4$4 and an extended component with $z\sim1.4$5 and $z\sim1.4$6. At 1300 $z\sim1.4$7m, the CARMA fit gives a compact component with $z\sim1.4$8 and an extended component with $z\sim1.4$9 and $\mu$0. The paper emphasizes that this morphology is unusual relative to previously studied SMGs, which are typically $\mu$1 in size.

A multiwavelength counterpart was identified near the CARMA/SMA position using Subaru/SuprimeCam $\mu$2 and $\mu$3, UKIDSS/UKIRT $\mu$4, Spitzer IRAC $\mu$5m, a Spitzer MIPS $\mu$6m upper limit, VLA 20 cm, and GMRT 50 cm. The optical/NIR source is red but visible, while the submm source is much brighter than expected from the optical light alone. The paper notes a possible positional offset between the optical peak and the submm/radio peaks of order $\mu$7–$\mu$8.

The preferred interpretation is a line-of-sight superposition. Hyperz applied to $\mu$9 yields

$S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$0

with confidence range $S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$1–$S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$2, and the optical/NIR SED shows a break near $S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$3m interpreted as a 4000 Å break. A GALAXEV fit at this redshift prefers a $S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$4 Myr instantaneous burst with stellar mass $S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$5 and extinction $S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$6. By contrast, the mm/submm/radio SED yields

$S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$7

which is more consistent with a typical SMG dust temperature of $S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$8 K and with the observed radio/submm colors.

Z-Spec line non-detections support this interpretation. The derived $S_{1100} = 37.28 \pm 0.65~{\rm mJy}.$9 upper limit on the line-to-continuum flux ratio is $30^{\prime\prime}$0–$30^{\prime\prime}$1 across the band. The paper argues that this is too low to easily detect mid-$30^{\prime\prime}$2 CO lines if the source were at $30^{\prime\prime}$3, but is reasonable if the source is at $30^{\prime\prime}$4, where higher-$30^{\prime\prime}$5 lines such as CO($30^{\prime\prime}$6) through CO($30^{\prime\prime}$7) would be observed and are often weak. It also reports $30^{\prime\prime}$8 $30^{\prime\prime}$9 and $0.6$0, helping argue that Orochi is probably at $0.6$1.

The foreground-background interpretation further motivates a lensing scenario. No obvious cluster lens is identified, but candidate foreground structures at $0.6$2 and $0.6$3 are noted, and the $0.6$4 galaxy is itself massive. The authors do not claim direct imaging evidence for strong lensing such as arcs or multiple images; accordingly, the lensing interpretation is suggestive rather than proven.

4. Infrared luminosity, gas reservoir, and astrophysical significance

Assuming that the submm-bright component lies at $0.6$5, the astronomical paper infers an apparent infrared luminosity of

$0.6$6

and a star formation rate of

$0.6$7

using Kennicutt’s relation and assuming that the infrared luminosity is dominated by star formation (Ikarashi et al., 2010).

If the system is instead forced to $0.6$8, the apparent values become $0.6$9 and $1.0$0, but this scenario requires an unusually cold dust temperature of $1.0$1 K, which the authors regard as unlikely. This makes the $1.0$2 interpretation the preferred one.

Using a dust-based estimate and a gas-to-dust ratio of 54, the paper gives

$1.0$3

for the favored high-redshift case, with depletion time

$1.0$4

That timescale is described as very short and comparable to intense starburst phases in the cores of local ULIRGs. By contrast, the $1.0$5 interpretation gives $1.0$6 yr, which the paper considers implausibly long for such an extreme submm-bright system.

The astronomical significance attributed to Orochi rests on several linked properties: extraordinary submm brightness, a resolved extended-plus-compact morphology, a likely lensed and optically dark high-redshift configuration, and an apparent luminosity and SFR in the regime of extreme starbursts. The paper further states that the source may be approaching the Eddington-limited starburst regime and that it illustrates how wide-area submm surveys can uncover rare, highly magnified galaxies.

5. Biomedical imaging scope and pretraining corpus

In biomedical imaging, Orochi is described as a unified processor for four low-level tasks: registration, fusion, restoration, and super-resolution. The motivating problem is the fragmented use of specialist models and plugins in tools such as ImageJ/Fiji and napari. The paper argues that fragmentation arises from three perspectives: workflows often require multiple steps; biomedical degradations are related across tasks; and biomedical images are huge and multi-dimensional, making multiple separate models inefficient (Dai et al., 26 Sep 2025).

The pretraining corpus consists of raw data from over 100 publicly available studies spanning 2D to 5D biomedical images with a total scale exceeding 100 TB. Because the datasets are too large to train on directly, the method converts raw images into patches or volumes. The sampling strategy is Random Multi-scale Sampling (RMS), defined by

$1.0$7

followed by random cropping,

$1.0$8

and assembly of the sampled set,

$1.0$9

The paper states that RMS captures different ROI sizes and resolutions, increases diversity across studies and modalities, makes storage and transmission more manageable, and supports both local training and streaming training on large datasets.

Orochi is also reported to work in zero-shot settings on unseen biomedical images after pretraining. That claim positions the method not only as a multi-task fine-tuning backbone but also as a general pretrained representation for low-level biomedical imaging.

6. Task-related pretraining, Mamba hierarchy, and fine-tuning

The central pretraining method is Task-related Joint-embedding Pre-Training (TJP), which replaces reliance on generic Masked Image Modelling with degradations designed to match biomedical low-level tasks. Its general self-supervised objective is written as

$\alpha=02^h18^m30.68^s$0

The degradation set includes masked images, deformed images, noisy images, and low-resolution images. For fusion, the paper uses dual-masking reconstructive fusion with $\alpha=02^h18^m30.68^s$1, $\alpha=02^h18^m30.68^s$2, and $\alpha=02^h18^m30.68^s$3. For restoration and super-resolution it uses spatially varying Gaussian downsampling, noisy downsampling, and multi-stage noise simulation. For registration it uses multi-scale smoothed Perlin noise deformation to generate anatomically plausible nonrigid deformation fields (Dai et al., 26 Sep 2025).

The architecture uses Mamba as the principal building block because of its linear computational complexity. The backbone is a Multi-head Hierarchy Mamba, described as a unified hierarchical Mamba encoder with a replaceable decoder. Design points explicitly listed are hierarchical feature extraction, patch merging, multi-scale encoding for 2D and 3D biomedical data, and task-specific tuning via the decoder or head. The appendix configuration gives embed_dim = 128, depths = (4,4,4,4), patch_size = 4, pat_merg_rf = 2, in_chans = 2, and out_indices = (0,1,2,3).

The fine-tuning framework has three tiers. Full fine-tunes 100% of parameters. Normal freezes the encoder and tunes only a regular convolution head, leaving about 10–30% of parameters trainable. Light freezes the encoder and trains only a depth-wise separable convolution head, with less than 5% of total parameters trainable. The paper explicitly frames this as a parameter-efficient fine-tuning strategy and notes that in data-limited settings, fewer trainable parameters can improve generalization by reducing overfitting. It further reports that 1–2% trainable parameters can work best in such settings.

For Orochi-B pretraining, the implementation details listed are batch size 12, learning rate 0.0005, weight decay 0.01, warmup ratio 0.1, maximum epoch 50, AdamW, and WarmupCosine, with A800 80G $\alpha=02^h18^m30.68^s$4 for local pretrain and H100 40G $\alpha=02^h18^m30.68^s$5 for streaming pretrain.

7. Benchmarks, ablations, and practical position

The biomedical paper compares Orochi against more than 30 baselines across restoration, super-resolution, registration, and fusion, using OASIS/Learn2Reg, VIFB, HBA, CARE, and additional appendix datasets such as IXI and Harvard SPECT-MRI or PET-MRI. The metrics reported are Dice, HD95, and SDlogJ for registration; Qabf, Qcv, and SSIM for fusion; and PSNR and SSIM for restoration and super-resolution (Dai et al., 26 Sep 2025).

Task	Orochi result	Comparator note
Restoration (CARE)	Light: 29.77 / 0.71 (XY), 29.98 / 0.72 (XZ)	Beats SwinIR, MambaIR, UniFMIR
Super-resolution (HBA)	Full: 35.33 / 0.95 and 31.93 / 0.89	At 8mm, exceeds InverseSR and VCM
Registration (OASIS/Learn2Reg)	Full: 83.62 Dice, 1.60 HD95, 0.11 SDlogJ	Higher Dice than TransMorph-L, MambaMorph, ConvexAdam, LapIRN
Fusion (VIFB CT-MRI)	Full: 0.41 Qabf, 2351.57 Qcv, 1.39 SSIM	Improves over BSAFusion in Qabf, Qcv, SSIM

The restoration results show an instructive pattern: on the CARE isotropic 3D volume restoration benchmark, Light outperforms Full and Normal, which the paper attributes to the small dataset size and the tendency of full fine-tuning to overfit. In super-resolution on HBA, Orochi Full yields 35.33 / 0.95 and 31.93 / 0.89, and the paper highlights gains over InverseSR and VCM at 8mm. In registration, Orochi Full reaches 83.62 Dice, 1.60 HD95, and 0.11 SDlogJ, while Orochi Light has the best SDlogJ at 0.06, suggesting smoother deformation at some cost in overlap accuracy. In fusion, Orochi Full reaches 0.41 Qabf, 2351.57 Qcv, and 1.39 SSIM, while Normal attains the highest SSIM at 1.45.

The ablation results are used to argue for task-related self-supervision rather than generic masked modeling. Reported values are Registration Dice 83.62 for Orochi versus 71.22 for MAE and 69.97 for I-JEPA; Fusion Qabf 0.41 versus 0.36 and 0.39; Restoration PSNR 29.88 versus 26.67 and 25.02; and Super-resolution PSNR 33.63 versus 29.17 and 28.81. A plausible implication is that the benefit comes not only from scale but from matching the pretraining corruption model to the downstream inverse problem.

Taken together, the biomedical results position Orochi as a single pretrained, task-aware, efficient model that can cover the principal low-level biomedical imaging workflows while remaining compatible with parameter-efficient fine-tuning. The astronomical results position Orochi as a rare, ultra-bright dusty starburst whose brightness, composite morphology, and likely lensing make it a distinctive object in early wide-field submillimeter surveys.