OccTENS: Dual Systems in OCT & Autonomous Driving

• OccTENS is an overloaded term for two separate systems: one for retinal OCT segmentation and another for 3D occupancy prediction in autonomous driving.
• In the medical imaging application, OccTENS employs a triple-encoder architecture that fuses spatial, spectral, and attentive features to improve segmentation accuracy.
• In autonomous driving, OccTENS utilizes Temporal Next-Scale Prediction with a transformer-based TensFormer to enable efficient long-horizon occupancy forecasting and pose controllability.

Searching arXiv for the cited OccTENS papers to ground the article. OccTENS is an overloaded research name that denotes two distinct 2025 arXiv systems in different subfields. In medical image analysis, OccTENS denotes an attentional triple-encoder network for retinal Optical Coherence Tomography (OCT) segmentation that combines spatial, spectral, and attention-based representations (Qi et al., 20 Mar 2025). In autonomous driving, OccTENS denotes a generative 3D occupancy world model based on Temporal Next-Scale Prediction (TENS) and a transformer called TensFormer, targeting long-horizon occupancy generation, computational efficiency, and controllable ego-motion conditioning (Jin et al., 4 Sep 2025). The two works are unrelated methodologically and application-wise, but both use the name to emphasize structured multi-branch or multi-scale modeling.

1. Terminological scope and disambiguation

The name “OccTENS” refers to two separate research artifacts rather than a single unified framework. One paper is titled “Attentional Triple-Encoder Network in Spatiospectral Domains for Medical Image Segmentation” and presents OccTENS as a retinal OCT segmentation model (Qi et al., 20 Mar 2025). A second paper is titled “OccTENS: 3D Occupancy World Model via Temporal Next-Scale Prediction” and presents OccTENS as a generative occupancy world model for autonomous driving (Jin et al., 4 Sep 2025).

Usage of “OccTENS”	Domain	Core formulation
Attentional Triple-Encoder Network in Spatiospectral Domains for Medical Image Segmentation	Retinal OCT segmentation	Triple-encoder fusion of CNN, FFC, and CNN-former branches
OccTENS: 3D Occupancy World Model via Temporal Next-Scale Prediction	Autonomous driving world modeling	Temporal Next-Scale Prediction with TensFormer and tokenized occupancy/pose

This naming collision is consequential for literature retrieval. In the medical-imaging paper, “OccTENS” is tied to OCT B-scan segmentation and spatiospectral fusion. In the autonomous-driving paper, “OccTENS” is tied to occupancy prediction, forecasting, and planning. A plausible implication is that citations and system descriptions should always be accompanied by the arXiv identifier to avoid ambiguity.

2. OccTENS in retinal OCT segmentation

In the medical-imaging usage, OccTENS is presented as a retinal OCT segmentation model whose core premise is that OCT images should be modeled as a coupled spatiospectral signal rather than only as 2D spatial textures or only in the frequency/spectral domain (Qi et al., 20 Mar 2025). The target task is segmentation of retinal OCT B-scans into anatomical layers and fluid regions on the Duke OCT dataset using Expert-1 annotations. The labels include ILM, NFL-IPL, INL, OPL, ONL-ISM, ISE, OS-RPE, and Fluid.

The motivation is explicitly tripartite. Spatial-domain CNN features are described as effective for local edge, texture, and shape modeling. Spectral-domain features extracted by Fourier processing are described as revealing periodicity, global structural patterns, and long-range dependencies that are hard to encode with convolutions alone. Attention is introduced to connect features across distant regions and across domains, enabling learning of global relationships and complementary cues. The paper argues that prior OCT segmentation methods tend to emphasize either spatial or spectral information, but not both jointly.

This formulation is particularly relevant to retinal OCT because boundaries between layers, fluid pockets, and thin pathological structures can be subtle locally while also exhibiting global regularities across slices and across scales. The paper therefore positions OccTENS against both purely spatial encoder-decoder designs and dual-encoder baselines, arguing that joint modeling of spatial and spectral dependencies, coupled with attention-based global relationship modeling, improves segmentation accuracy.

3. Triple-encoder architecture and feature fusion

The retinal OccTENS architecture is organized into two stages. Stage 1 contains three parallel encoders that process the same OCT input: a CNN branch for spatial/local representation, an FFC branch for spectral/frequency-domain representation, and a CNN-former branch for global relationship modeling via attention and convolution (Qi et al., 20 Mar 2025). The paper states that each of the CNN and FFC branches has four blocks, and the CNN-former also has four sub-blocks.

The CNN-based spatial branch is the local-detail extraction path. It processes OCT image input, uses Dynamic ReLU, and uses two MLP layers, producing feature maps that feed into the CNN-former path. Its output is denoted $\tilde{f}^{cnn}$. The FFC branch is the spectral encoder. Its intended workflow is to transform features into the frequency domain using Fourier transform, perform convolution-like processing with global receptive fields in that domain, and transform back to the spatial domain or return frequency-aware feature maps. Its output is denoted $\tilde{f}^{ffc}$.

The third encoder, the CNN-former, combines a CNN sub-block using Dynamic ReLU and MLP layers with a Former sub-block using multi-head attention and feed-forward networks. The branch also includes a lightweight cross-attention module mapping CNN features into attention tokens and another lightweight cross-attention module mapping refined attention tokens back to CNN features. This is described as being adapted from the Mobile-Former style of bridging convolution and transformer representations. Its output is denoted $\tilde{f}^{former}$.

Stage 2 performs fusion and decoding. Features from the three encoders are concatenated and refined with cross-attention modules and convolutional decoding. The decoder iteratively fuses multi-level features with four cross-attention modules while preserving context and producing the final segmentation mask. The key fusion equation is given as

$\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$

which combines cross-attention between CNN and FFC features and between FFC and former features.

The paper characterizes the fusion strategy as a combination of concatenation for multi-branch feature aggregation, cross-attention for relation modeling, and addition for combining attention outputs. This suggests that semantic alignment across branches is treated as a central problem, not merely channel aggregation.

4. Optimization, training protocol, and segmentation results

The retinal OccTENS paper uses a combined Dice and cross-entropy objective

$\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$

with $\lambda_{Dice}=1$ and $\lambda_{CE}=1$ during training (Qi et al., 20 Mar 2025). The training setup is reported as follows: Duke OCT dataset; 10 patients total; training subjects 1–6; validation subjects 7–8; testing subjects 9–10; Expert-1 as annotation reference; input resized to $224 \times 224$; Adam optimizer; learning rate $5 \times 10^{-4}$; weight decay $1 \times 10^{-4}$; and 100 epochs.

The principal evaluation metric is Dice score per layer and mean Dice across all layers. The main reported result is a mean Dice of $\tilde{f}^{ffc}$0, improving over the Y-Net baseline’s $\tilde{f}^{ffc}$1. The comparison table in the paper lists RelayNet at $\tilde{f}^{ffc}$2, Language at $\tilde{f}^{ffc}$3, Alignment at $\tilde{f}^{ffc}$4, U-Net at $\tilde{f}^{ffc}$5, Y-Net at $\tilde{f}^{ffc}$6, and OccTENS at $\tilde{f}^{ffc}$7. The improvement over Y-Net is explicitly computed as $\tilde{f}^{ffc}$8, i.e., a $\tilde{f}^{ffc}$9 percentage point increase in mean Dice.

Per-class highlights include Fluid Dice $\tilde{f}^{former}$0 and ILM Dice $\tilde{f}^{former}$1. The paper interprets this as evidence of strong performance on difficult high-variance regions such as fluid while also improving boundary layers. In practical terms, the method is described as useful for retinal layer segmentation, intraretinal fluid segmentation, clinical OCT analysis, and improving downstream diagnosis and treatment monitoring.

The ablation study focuses on cross-attention placement. Removing cross-attention in the Stage 1 encoder yields a mean Dice of $\tilde{f}^{former}$2. Removing decoder cross-attention yields a mean Dice of $\tilde{f}^{former}$3. Both are below the full model’s $\tilde{f}^{former}$4, supporting the paper’s claim that encoder-side cross-attention improves feature fusion and decoder-side cross-attention preserves multi-scale relational structure. The stated future direction is validation on other OCT datasets. A plausible limitation, explicitly inferred in the source material, is that evaluation is currently limited to the Duke OCT dataset with only 10 patients, so broader validation is needed to establish generalizability across devices, disease types, and acquisition settings.

5. OccTENS as a 3D occupancy world model

In the autonomous-driving usage, OccTENS is a generative 3D occupancy world model that reframes occupancy as a world representation to be predictively generated over time from historical observations, enabling downstream planning and simulation (Jin et al., 4 Sep 2025). The paper identifies three limitations in recent occupancy world models based on vanilla autoregressive next-token prediction over flattened occupancy tokens: inefficiency, temporal degradation in long-term generation, and poor controllability with respect to desired camera pose or trajectory.

OccTENS addresses these issues by reformulating occupancy world modeling as Temporal Next-Scale Prediction, abbreviated TENS. The key decomposition is between temporal scene-by-scene prediction and spatial scale-by-scale generation. The paper argues that occupancy has stronger geometric structure than standard visual generation tasks, and that naïve frame-wise autoregression conflates inter-frame temporal causality with intra-frame bidirectional spatial dependencies. TENS is introduced specifically to separate these axes.

The paper first states the vanilla autoregressive formulation over a flattened occupancy feature map of resolution $\tilde{f}^{former}$5 as

$\tilde{f}^{former}$6

It then notes that if next-scale prediction is used directly, the token sequence becomes much longer, with the total number of tokens for $\tilde{f}^{former}$7 scales approximately

$\tilde{f}^{former}$8

This motivates the paper’s structured temporal-and-scale decomposition.

6. Tokenization, TensFormer, controllability, and empirical results

The autonomous-driving OccTENS has two major components: tokenizers for occupancy and ego motion, and TensFormer as the generative world model (Jin et al., 4 Sep 2025). The scene tokenizer converts a 3D occupancy scene $\tilde{f}^{former}$9 into a BEV latent map

$\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$0

then uses a multi-scale quantizer to produce discrete token maps

$\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$1

The motivation given is that occupancy scenes have interdependent spatial structure, so a single local codebook tokenization can lose global context.

The motion tokenizer discretizes relative ego motion in $\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$2, $\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$3, and orientation $\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$4, ignoring $\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$5 because it is usually negligible for driving. Using vocabularies $\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$6, $\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$7, and $\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$8, motion is mapped to a token by Cartesian composition:

$\begin{aligned} & \operatorname{Attn}=\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{cnn}, \tilde{f}^{ffc}\right)+\operatorname{Attn}_{\text {cross }\left(\tilde{f}^{ffc}, \tilde{f}^{former}\right) \end{aligned} \tag{1}$9

where $\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$0 is an embedding layer. This means pose is incorporated as part of sequence modeling rather than only as an external conditioning vector.

TensFormer explicitly separates temporal and spatial reasoning. For temporal scene-by-scene prediction, the model uses

$\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$1

It adds position, scale, and time embeddings, then decomposes attention into scale-wise temporal causal attention and frame-wise spatial attention. For scale $\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$2 at time $\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$3, tokens can attend to all previous frames and lower scales within the current frame:

$\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$4

Spatial scale-by-scale generation is then modeled as

$\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$5

A major contribution is the holistic pose aggregation strategy, in which motion is treated as the $\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$6-th scale token:

$\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$7

The autoregressive factorization then becomes

$\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$8

The paper describes this as holistic sequence modeling of occupancy and ego-motion, enabling both pose controllability and motion planning.

The tokenizer is trained with

$\mathcal{L}_{\text {total }=\lambda_{\text {Dice } \mathcal{L}_{\text {Dice }+\lambda_{C E} \mathcal{L}_{C E}, \tag{2}$9

and the world model with

$\lambda_{Dice}=1$0

The evaluation uses nuScenes with Occ3D occupancy annotation, a 700/150 train/val split, about 20 seconds per sequence, keyframes at 2 Hz, and a default setting of 2 seconds of history to forecast 3 seconds into the future.

Quantitatively, for 4D occupancy forecasting with occupancy input, OccWorld-O reports Avg mIoU $\lambda_{Dice}=1$1 and Avg IoU $\lambda_{Dice}=1$2, OccLLaMA-O reports $\lambda_{Dice}=1$3 and $\lambda_{Dice}=1$4, and OccTENS-O reports $\lambda_{Dice}=1$5 and $\lambda_{Dice}=1$6. With camera input, OccWorld-F reports $\lambda_{Dice}=1$7 and $\lambda_{Dice}=1$8, OccLLaMA-F reports $\lambda_{Dice}=1$9 and $\lambda_{CE}=1$0, and OccTENS-F reports $\lambda_{CE}=1$1 and $\lambda_{CE}=1$2. For planning, OccWorld reports L2 $\lambda_{CE}=1$3 and collision $\lambda_{CE}=1$4, OccLlama reports L2 $\lambda_{CE}=1$5 and collision $\lambda_{CE}=1$6, and OccTENS reports L2 $\lambda_{CE}=1$7 and collision $\lambda_{CE}=1$8. The efficiency ablation reports latency of $\lambda_{CE}=1$9 s for Ours-2-scales, $224 \times 224$0 s for Ours-4-scales, $224 \times 224$1 s for Ours-6-scales, and $224 \times 224$2 s for Ours-8-scales, while OccWorld reports $224 \times 224$3 s and OccSora reports approximately $224 \times 224$4 s. The authors choose 6 scales as the best trade-off. The stated implementation details include tokenization downsampling factor 8, codebook size 4096, codebook dimension 128, multi-scale setting with 6 scales and $224 \times 224$5, 4 layers per block, hidden dimension 128, 4 attention heads, and $224 \times 224$6.

7. Comparative interpretation and significance

The two OccTENS systems share a formal design intuition: both reject a single-stream formulation in favor of structured decomposition. In retinal OCT segmentation, the decomposition is across spatial, spectral, and attention-based encoders followed by cross-attention fusion (Qi et al., 20 Mar 2025). In occupancy world modeling, the decomposition is across temporal scene evolution, spatial scale refinement, and unified occupancy-pose tokenization (Jin et al., 4 Sep 2025).

Despite this superficial parallel, the methods operate on different objects, objectives, and evaluation regimes. The retinal OccTENS consumes OCT B-scans resized to $224 \times 224$7, optimizes Dice plus cross-entropy, and reports mean Dice and per-layer Dice on the Duke OCT dataset. The autonomous-driving OccTENS consumes tokenized occupancy scenes and ego-motion, optimizes token prediction losses, and reports mIoU, IoU, L2 error, collision rate, and latency on nuScenes/Occ3D. The name therefore does not denote a transferable algorithmic family in the narrow sense.

A plausible interpretation is that both papers use “OccTENS” to signal structured occupancy or OCT modeling through explicit factorization rather than monolithic autoregression or single-domain encoding. In the medical paper, the factorization addresses local detail, Fourier-domain regularity, and global relational reasoning. In the driving paper, it addresses long-horizon temporal causality, coarse-to-fine spatial generation, and pose controllability. For literature indexing, benchmarking, and citation practice, the distinction between OccTENS (Qi et al., 20 Mar 2025) and OccTENS (Jin et al., 4 Sep 2025) is therefore essential.