XOCT: Dual-Use in Excitonic Devices & OCT Imaging

• XOCT is a term defining both an optically controlled excitonic transistor, which uses crossed-ramp architectures to steer indirect exciton flow, and a deep learning framework for OCT-to-OCTA translation.
• The excitonic device employs a drift–diffusion model with optical gating to achieve up to 100× photoluminescence enhancement and precise routing at cryogenic temperatures.
• The OCT-to-OCTA method integrates Cross-Dimensional Supervision and Multi-Scale Feature Fusion to generate high-fidelity angiographic projections, outperforming several state-of-the-art models.

Searching arXiv for papers on "XOCT" and related usage to ground the article in current literature. XOCT is an acronym used in two unrelated technical contexts represented in arXiv literature: the optically controlled excitonic transistor, a planar coupled-quantum-well device for all-optical switching and routing of indirect exciton fluxes, and XOCT, a deep learning framework for enhancing Optical Coherence Tomography (OCT) to Optical Coherence Tomography Angiography (OCTA) translation via Cross-Dimensional Supervision and Multi-Scale Feature Fusion [(Andreakou et al., 2013); (Khosravi et al., 9 Sep 2025)]. The shared acronym conceals a sharp disciplinary divergence: one usage belongs to semiconductor excitonics at cryogenic temperature, while the other belongs to retinal image synthesis and computational ophthalmic imaging. This suggests that XOCT is best interpreted contextually rather than as a single canonical concept.

1. Terminological scope and domain separation

In the literature considered here, XOCT denotes either an optically controlled excitonic transistor or a layer-aware OCT-to-OCTA translation model. The first usage is associated with a crossed-ramp excitonic device in a GaAs/AlGaAs coupled-quantum-well heterostructure that demonstrates experimental proof of principle for all-optical excitonic transistors and all-optical excitonic routers. The second usage is associated with a 3D encoder–decoder framework that integrates Cross-Dimensional Supervision (CDS) with a Multi-Scale Feature Fusion (MSFF) network for retinal vascular reconstruction from OCT volumes [(Andreakou et al., 2013); (Khosravi et al., 9 Sep 2025)].

XOCT usage	Research domain	Defining elements
Optically controlled excitonic transistor	Excitonic transport and semiconductor devices	Crossed ramps, indirect excitons, optical source/gate/drain
XOCT for OCT-to-OCTA translation	Medical image translation and ophthalmic AI	CDS, MSFF, 3D encoder–decoder, OCTA-500

A common misconception is to treat XOCT as a standardized acronym with a single meaning. The available papers instead show an acronym reused across distinct problem settings. A plausible implication is that citation context is essential whenever XOCT appears in interdisciplinary discussions.

2. Optically controlled excitonic transistor: device architecture and materials

The optically controlled excitonic transistor is built on a GaAs/AlGaAs coupled-quantum-well (CQW) heterostructure grown by molecular-beam epitaxy. Its core consists of two GaAs quantum wells, each approximately 8 nm thick, separated by a 4 nm $\mathrm{Al}_{0.33}\mathrm{Ga}_{0.67}\mathrm{As}$ barrier. The CQW is embedded between a uniform $n^+$-GaAs bottom electrode and a patterned semitransparent top electrode made of 30 nm Ti/Pt. The top electrode is shaped into two narrow ramps that cross at a right angle, each ramp being a wedge whose width narrows in the direction of exciton flow and thereby creates a linear in-plane potential gradient for indirect excitons. Wider flat channels of constant electrode width lie on both sides of each ramp, and the two ramps cross at a central junction to form a four-arm geometry (Andreakou et al., 2013).

The active quasiparticles are indirect excitons, consisting of an electron and hole confined in separate quantum wells. Their built-in dipole moment is given as $p \approx ed$ with $d \approx 12\,\mathrm{nm}$, so their energy shifts by $edF_z$ under an applied vertical field $F_z$. The structure provides strong confinement through a typical conduction-band offset $\Delta E_c \approx 200\,\mathrm{meV}$ and valence-band offset $\Delta E_v \approx 100\,\mathrm{meV}$. The exciton binding energy for direct excitons is approximately 10 meV, while for indirect excitons it is reduced to a few meV. The corresponding radiative lifetimes are $\tau_0 \sim 10$–$100\,\mathrm{ns}$, orders of magnitude longer than for direct excitons, allowing diffusion over tens of microns (Andreakou et al., 2013).

These material and geometric choices determine the operating regime. The long lifetime is not merely a materials detail; it is the condition that makes transport over the crossed-ramp geometry experimentally accessible.

3. Excitonic switching, routing, and transport model

Operation relies on optical source, gate, and drain beams. A tightly focused source laser with $n^+$0 creates excitons at the input port of one ramp, labeled S. These excitons drift and diffuse down the energy ramp toward the cross-junction. A second, weaker gate laser, also at $n^+$1, is focused either on one arm of the crossing region or at a separate gate location. The gate beam locally creates excitons that screen the underlying disorder via repulsive dipole–dipole interactions, heat the exciton gas, and partially fill the potential valley, thereby permitting or blocking the passage of source-generated excitons. The drain is not a separate contact; it is the photoluminescence collected from the downstream region of the chosen ramp arm beyond the crossing point (Andreakou et al., 2013).

The device exhibits distinct OFF and ON states. In the OFF state, with the source beam only, exciton transport is arrested upstream of the junction by disorder and by the unmodified potential gradient, so the drain photoluminescence is weak. In the ON state, with both source and gate beams, gate-induced excitons screen disorder and locally raise the effective exciton temperature and lifetime, enabling source excitons to surmount residual barriers and arrive at the drain. The resulting photoluminescence increases by up to two orders of magnitude. Routing is realized by placing the gate spot on one ramp arm or the other, directing the source exciton flux into the corresponding drain arm (Andreakou et al., 2013).

The electrostatic mechanism is encoded in the exciton potential landscape $n^+$2. A dc bias $n^+$3 applied between the patterned top electrode and the uniform bottom electrode creates a nominally uniform vertical field $n^+$4 under the wide parts of the top electrode. Where the top electrode narrows, field lines fringe outward, the local $n^+$5 is reduced, and the exciton potential energy $n^+$6 is raised. The spatially varying electrode width therefore imprints a linear in-plane gradient along each ramp. The steady-state transport model is a drift–diffusion–generation–recombination equation,

$n^+$7

where $n^+$8 is the exciton density, $n^+$9 the diffusion coefficient, $p \approx ed$0 the mobility, $p \approx ed$1 the in-plane force, $p \approx ed$2 the effective optical lifetime, and $p \approx ed$3 the local generation rate from source and gate lasers. The potential energy is written as

$p \approx ed$4

where $p \approx ed$5 is the static band-edge profile set by the electrode shape and $p \approx ed$6 is the dipole moment (Andreakou et al., 2013).

This formulation makes clear that the transistor action is not based on charge injection through conventional metallic terminals. Instead, it is governed by optically generated indirect excitons evolving in a spatially engineered potential.

4. Excitonic performance metrics, operating regime, and applications

Performance is expressed through on/off contrast and excitonic gain. The on/off contrast ratio is defined as the ratio of drain emission in the ON state to drain emission in the OFF state,

$p \approx ed$7

while the excitonic gain is the ratio of the ON-state drain signal to the gate-only signal under identical gate power,

$p \approx ed$8

As the gate power $p \approx ed$9 increases from $d \approx 12\,\mathrm{nm}$0 to approximately $d \approx 12\,\mathrm{nm}$1, with source power $d \approx 12\,\mathrm{nm}$2 fixed at $d \approx 12\,\mathrm{nm}$3, the integrated drain emission rises by up to $d \approx 12\,\mathrm{nm}$4. The switching threshold occurs at $d \approx 12\,\mathrm{nm}$5–$d \approx 12\,\mathrm{nm}$6. The maximum on/off contrast reaches $d \approx 12\,\mathrm{nm}$7, and the gain reaches $d \approx 12\,\mathrm{nm}$8 (Andreakou et al., 2013).

The reported operating regime is strongly constrained by lifetime, transport length, temperature, and bias. Because $d \approx 12\,\mathrm{nm}$9–$edF_z$0 and exciton diffusion times over $edF_z$1 are tens of ns, the intrinsic switching time is projected to be on the order of $edF_z$2–$edF_z$3. The optical spot size, with FWHM $edF_z$4, and exciton transport lengths of $edF_z$5 set a routing resolution of a few microns. Measurements were taken at $edF_z$6, and the device ceases to function above $edF_z$7, where exciton binding is thermally quenched. The bias $edF_z$8–$edF_z$9 sets $F_z$0; lower bias reduces ramp height and suppresses transport, while higher bias increases nonradiative leakage (Andreakou et al., 2013).

The applications discussed include excitonic logic arrays, routers, and reconfigurable interconnects. The crossed-ramp architecture is described as naturally extensible to fan-out and reconfigurable interconnects, and multiple XOCTs may be combined into multiplexers, demultiplexers, or all-optical neural networks operating at cryogenic temperatures. The paper also contrasts optical and electrical gating: optical gating requires no electrical contacts for local control and can be dynamically reconfigured on sub-$F_z$1 time scales, whereas electrical gates suffer from RC delays and Joule heating (Andreakou et al., 2013).

5. XOCT for OCT-to-OCTA translation: problem setting and architecture

In retinal imaging, XOCT denotes a deep learning framework designed to translate standard OCT volumes into OCTA volumes and derived en-face projections. The motivating problem is twofold. OCTA acquisition is highly sensitive to patient motion and requires hardware or software upgrades to standard OCT systems, increasing clinical cost and limiting accessibility. At the same time, the retina is a multi-layered structure in which each lamina exhibits distinct vascular patterns and imaging characteristics, and naïve volumetric translation from OCT to OCTA often fails to reconstruct thin capillary plexuses or maintain vascular continuity across layers, producing fragmented or blurred vessels that undermine clinical utility in diseases such as diabetic retinopathy and age-related macular degeneration (Khosravi et al., 9 Sep 2025).

The proposed framework introduces two modules built on top of a 3D encoder–decoder architecture: Cross-Dimensional Supervision (CDS) and Multi-Scale Feature Fusion (MSFF). CDS exploits segmentation maps of retinal layers during training to generate layer-specific 2D en-face OCTA projections. Given a predicted OCTA volume $F_z$2 and a binary segmentation map $F_z$3 for layer $F_z$4, the layer-wise projection is computed by segmentation-weighted averaging along the depth axis:

$F_z$5

This ensures that only voxels belonging to layer $F_z$6 contribute to its projection. The network is then guided to match each $F_z$7 to its ground-truth counterpart $F_z$8 through a composite CDS loss consisting of an $F_z$9 term, an adversarial term, and a perceptual term computed via a pre-trained VGG19 network (Khosravi et al., 9 Sep 2025).

The significance of CDS lies in its explicit alignment between 3D volumetric prediction and 2D clinically salient projections. By supervising each layer individually, the method compels the encoder–decoder to learn distinct feature subspaces that respect layer-specific vessel topology and contrast properties.

6. Multi-scale fusion, optimization, and empirical results

Where CDS enforces layer-aware supervision, MSFF is designed to capture the wide dynamic range of vessel calibers, from fine capillaries to larger arterioles, within a single hierarchy. MSFF extracts multi-scale representations in three parallel branches: isotropic $\Delta E_c \approx 200\,\mathrm{meV}$0 convolutions for balanced local context, anisotropic kernels $\Delta E_c \approx 200\,\mathrm{meV}$1 to accentuate elongated vessel patterns, and a depth-wise large $\Delta E_c \approx 200\,\mathrm{meV}$2 convolution to broaden the receptive field and capture global vessel continuity. The outputs are projected to a common channel dimension through $\Delta E_c \approx 200\,\mathrm{meV}$3 convolutions, concatenated, fused by a point-wise convolution, modulated by a two-layer channel-attention transform based on global average pooling, and combined with a residual connection to preserve low-level details and facilitate gradient flow (Khosravi et al., 9 Sep 2025).

Training uses the public OCTA-500 dataset, comprising 500 paired OCT/OCTA volumes with expert retinal layer segmentations. The data are split into OCTA-3M with $\Delta E_c \approx 200\,\mathrm{meV}$4 voxels and 140/20/40 train/val/test, and OCTA-6M with $\Delta E_c \approx 200\,\mathrm{meV}$5 voxels and 200/30/70. For each sample, the raw OCT volume $\Delta E_c \approx 200\,\mathrm{meV}$6 is input to the 3D generator $\Delta E_c \approx 200\,\mathrm{meV}$7 and yields $\Delta E_c \approx 200\,\mathrm{meV}$8. Optimization uses the combined loss

$\Delta E_c \approx 200\,\mathrm{meV}$9

with

$\Delta E_v \approx 100\,\mathrm{meV}$0

where $\Delta E_v \approx 100\,\mathrm{meV}$1. All adversarial terms carry weight $\Delta E_v \approx 100\,\mathrm{meV}$2, perceptual weights $\Delta E_v \approx 100\,\mathrm{meV}$3 are set to $\Delta E_v \approx 100\,\mathrm{meV}$4 by grid search, and training uses the Adam optimizer, learning rate $\Delta E_v \approx 100\,\mathrm{meV}$5, batch size 1, and 300 epochs without learning-rate decay (Khosravi et al., 9 Sep 2025).

Quantitative evaluation employs MAE, PSNR, SSIM, and Perceptual Discrepancy. XOCT is reported to consistently outperform BBDM, Pix2Pix, MultiGAN, BBDM3D*, Pix2Pix3D, and TransPro. On the OCTA-3M en-face full-volume projection, XOCT achieves MAE $\Delta E_v \approx 100\,\mathrm{meV}$6, PSNR $\Delta E_v \approx 100\,\mathrm{meV}$7, and SSIM $\Delta E_v \approx 100\,\mathrm{meV}$8, compared with TransPro at 19.54, 20.14, and 0.580. In layer-specific projections, the SSIM on the ILM–OPL plane rises from 0.509 to 0.577. The ablation study reports that adding CDS to a Pix2Pix3D backbone raises projection SSIM from 0.556 to 0.600, MSFF alone boosts 3D SSIM from 0.885 to 0.893, and combining both modules yields the strongest overall outcome (Khosravi et al., 9 Sep 2025).

The clinical framing is explicit: the method is intended to remove the barrier of specialized OCTA hardware by enabling high-fidelity angiograms from standard OCT scans, while preserving diagnostic cues such as capillary dropout in diabetic retinopathy and neovascular tufts in wet AMD. Future work is stated to focus on robust domain adaptation to different OCT devices, further enhancement of minute vessel reconstruction, and optimization for real-time deployment, potentially integrating multi-modal retinal data (Khosravi et al., 9 Sep 2025).

7. Comparative significance and contextual interpretation

The two XOCT usages share an emphasis on optical control or optical inference, but they solve different classes of problems. The excitonic XOCT uses optical beams to regulate the motion of indirect excitons in a fabricated semiconductor potential landscape, with switching behavior determined by disorder screening, heating of the exciton gas, and potential-valley filling. The retinal XOCT uses a 3D generator trained with layer-specific 2D supervision and multi-scale feature fusion to infer angiographic structure from structural OCT data [(Andreakou et al., 2013); (Khosravi et al., 9 Sep 2025)].

This contrast helps clarify two further points. First, in the excitonic setting, the drain is a photoluminescence readout region rather than an electronic terminal in the conventional transistor sense. Second, in the retinal-imaging setting, XOCT is not an OCTA acquisition device but a translation framework that operates on paired OCT/OCTA training data. A plausible implication is that the acronym’s reuse reflects local naming logic within each field rather than any substantive methodological lineage between the two.

Viewed together, the two usages illustrate how the same acronym can attach to sharply different research programs: one centered on cryogenic excitonic interconnects and all-optical routing, the other on layer-aware volumetric-to-angiographic synthesis for ophthalmic diagnosis. The terminological overlap is therefore incidental, whereas the technical content is domain-specific and should be interpreted through the corresponding arXiv record.