Triplet-UNet Architecture

• Triplet-UNet Architecture is a family of U-Net designs that employs three coordinated streams to separate and process complementary evidence for improved image segmentation.
• It is applied across diverse domains including brain tumor, skin lesion, nuclei instance segmentation, and remote-sensing change detection.
• Key fusion methods—such as probability averaging, progressive feature aggregation, attention mechanisms, and cascade control—optimize prediction accuracy.

Searching arXiv for the cited Triplet-/Triple-/Triplet U-Net papers to ground the article in current records. Searching for (Tang et al., 2024) and related Triplet/Triple U-Net variants. Using arXiv search to retrieve the relevant records. Triplet-UNet Architecture denotes a family of U-Net-derived designs in which three coordinated subnetworks, views, or feature streams are used within a single segmentation or change-detection pipeline. In the arXiv literature, the name is applied to several non-isomorphic designs rather than to one canonical topology: a triplanar ensemble of 2D U-Nets for brain tumour segmentation (Sundaresan et al., 2021), a three-branch nuclei instance segmenter built from RGB, Hematoxylin, and segmentation branches (Ahmed et al., 2024), a region-of-interest-focused three-stage skin-lesion segmenter (Liu et al., 2023), a triplet-encoder remote-sensing model with explicit change-image processing (Zhong et al., 2023), a tree-structured cascade of three 3D U-Nets for biologically informed glioma segmentation (Beers et al., 2017), and a recent TM-UNet that combines residual VSS Blocks with a Triplet SSM module for medical image segmentation (Tang et al., 2024). The common architectural invariant is therefore the use of three coordinated computational paths, while the meaning of “triplet” depends on whether the decomposition is by plane, modality, time phase, or stage.

1. Nomenclature and scope

The published record uses closely related labels—“Triplet U-Net,” “Triple U-Net,” “Triple-UNet,” “T-UNet,” and “TM-UNet”—for architectures that share a three-part organization but differ substantially in their internal operators, fusion rules, and supervision strategies.

Variant	Triplet organization	Task
Triplanar ensemble of U-Nets (Sundaresan et al., 2021)	Independent axial, sagittal, and coronal 2D U-Nets, plus an independent tumour-core prediction module	Brain tumour segmentation
Triple U-Net (Ahmed et al., 2024)	RGB branch, Hematoxylin branch, and Segmentation branch with PDFA	Nuclei instance segmentation
Triple-UNet (Liu et al., 2023)	Three UNet architectures linked by ROIE and stage-wise refinement	Skin lesion segmentation
T-UNet (Zhong et al., 2023)	Three-branch encoder over pre-change, post-change, and difference images	High-resolution remote-sensing change detection
Sequential 3D U-Nets (Beers et al., 2017)	U-Net_WT, U-Net_ET, and U-Net_TC in a tree structure	Glioma tissue segmentation
TM-UNet (Tang et al., 2024)	Residual VSS Blocks and Triplet SSM	Medical image segmentation

This variability is important for interpretation. In one line of work, “triplet” means three orthogonal views of the same volume; in another, it means three synchronized branches with heterogeneous inputs; in another, it denotes a three-stage cascade. The term therefore refers to an architectural family rather than a standardized blueprint.

2. Principal architectural forms

One major form is the triplanar ensemble. In brain MR segmentation, the network consists of three independent 2D U-Nets of depth $=3$ down-up levels, operating on axial, sagittal, and coronal slices extracted from four co-registered MR modalities. The three branches share the same high-level encoder-decoder topology but use different kernel sizes: the axial branch uses $3\times 3$ convolutions, whereas the sagittal and coronal branches use $5\times 5$ convolutions. Each branch terminates in a $1\times 1$ convolution to $C=4$ channels followed by softmax, and a fourth axial network predicts tumour core for post-processing refinement (Sundaresan et al., 2021).

A second form is the three-branch parallel U-Net with progressive feature aggregation. In nuclei instance segmentation on CryoNuSeg, each of the RGB, Hematoxylin, and Segmentation branches follows a 5-level U-Net with encoder blocks of the form “Conv$3\times 3$\rightarrow$ReLU$\rightarrow$Conv$3\times 3$\rightarrow$ReLU,” $3\times 3$0 max-pooling in the encoder, and $3\times 3$1 transposed convolution in the decoder. The Segmentation branch does not operate independently; at each encoder and decoder level it receives progressively fused features from the RGB and Hematoxylin branches through the PDFA module (Ahmed et al., 2024).

A third form is the three-stage cascade with explicit region emphasis. For skin lesion segmentation, all three subnetworks use a “light” UNet-style encoder-decoder backbone built from depthwise-separable convolutions, BatchNorm, ReLU, and a channel-attention module, with filter counts $3\times 3$2. UNet 1 produces a score map, UNet 2 receives an ROIE-enhanced image and decoder-side feature concatenations from UNet 1, and UNet 3 takes the original image masked by the second score map and concatenates encoder features from all three subnetworks in its decoder (Liu et al., 2023).

A fourth form is the triplet encoder for multi-temporal change detection. T-UNet processes $3\times 3$3, $3\times 3$4, and the point-wise difference $3\times 3$5 using three VGG16-style encoder branches. The $3\times 3$6 and $3\times 3$7 branches share ImageNet-pretrained VGG16 weights, while the difference branch uses distinct weights. After each convolutional module in the difference branch, MBSSCA fuses features from the three branches; the decoder then combines a convolutional module, SAM, $3\times 3$8 transposed convolution, and CAM-conditioned skip fusion (Zhong et al., 2023).

A fifth form is the tree-structured 3D cascade. In biologically informed brain-tumour segmentation, U-Net$3\times 3$9 predicts the whole-tumour region first. Its thresholded output is then concatenated with the original four-channel MR patch and passed to separate U-Net$5\times 5$0 and U-Net$5\times 5$1 models. Each module uses the same base 3D U-Net: $5\times 5$2 input patches, four resolution levels, two $5\times 5$3 convolutions plus batch normalization and ReLU per level, $5\times 5$4 max-pooling on the down path, transposed convolutions on the up path, and a final $5\times 5$5 convolution with sigmoid (Beers et al., 2017).

TM-UNet extends the family beyond pure CNN-based and VGG-style realizations. Its abstract states that the model leverages residual VSS Blocks to extract intensive contextual features, while Triplet SSM is employed to fuse features across spatial and channel dimensions (Tang et al., 2024).

3. Cross-branch fusion and information flow

In the triplanar ensemble, fusion is performed at the probability level rather than through trainable cross-branch interaction. After slice-wise inference, the probability maps are reassembled into three $5\times 5$6 volumes and averaged: $5\times 5$7 The final label is then

$5\times 5$8

No trainable weighting or late-fusion layer is used (Sundaresan et al., 2021).

In the skin-lesion cascade, information transfer is both image-level and feature-level. The ROIE module uses the first score map $5\times 5$9 to produce an enhanced image

$1\times 1$0

In practice, $1\times 1$1, so $1\times 1$2. UNet 2 then combines its own features with encoder features from UNet 1, and UNet 3 takes $1\times 1$3 while concatenating encoder features from UNet 1, UNet 2, and UNet 3 (Liu et al., 2023).

In the nuclei model, the central fusion operator is PDFA. Instead of naïve concatenation of all three branches’ feature maps at each resolution, PDFA merges them in a small cascade of $1\times 1$4 encoder-side or $1\times 1$5 decoder-side $1\times 1$6 Conv+ReLU layers. With feature tensors $1\times 1$7, $1\times 1$8, and $1\times 1$9, the fused representation is written as

$C=4$0

with intermediate concatenation after each aggregation step (Ahmed et al., 2024).

In T-UNet, fusion is attention-mediated. At level $C=4$1, MBSSCA first forms a channel-attended tensor

$C=4$2

then computes spatial attention weights from both $C=4$3 and $C=4$4, combines them as

$C=4$5

and outputs

$C=4$6

This arrangement is intended to suppress pseudo-changes in $C=4$7 and enhance true changes via spectral and spatial attention across all three branches (Zhong et al., 2023).

In the 3D cascade, the triplet interaction is sequential rather than simultaneous. The whole-tumour prediction $C=4$8 is thresholded to a mask $C=4$9, and the inputs to the ET and TC networks become $3\times 3$0 and $3\times 3$1. This makes the whole-tumour mask an explicit conditioning signal for the downstream tumour-subregion predictors (Beers et al., 2017).

4. Objectives, supervision, and optimization

Triplet-UNet variants also differ in how supervision is distributed across branches. In the triplanar brain-tumour ensemble, both the three-class branches and the tumour-core branch use a combined cross-entropy plus Dice loss per mini-batch. Training uses Adam with $3\times 3$2, batch size $3\times 3$3, an initial learning rate of $3\times 3$4 decayed by $3\times 3$5 every $3\times 3$6 epochs until $3\times 3$7, and online augmentation with translation, rotation, and Gaussian noise. The three triplanar branches are trained independently, as is the TC branch (Sundaresan et al., 2021).

In the nuclei instance segmenter, supervision is branch-specific. The RGB branch uses binary cross-entropy, the Hematoxylin branch uses Soft-Dice to emphasize contour learning, and the Segmentation branch uses both cross-entropy and Soft-Dice. The full objective is

$3\times 3$8

with $3\times 3$9's chosen to balance the four objectives; the description gives the example that all are set to $\rightarrow$0 (Ahmed et al., 2024).

In the skin-lesion Triple-UNet, each stage is supervised directly against the ground-truth mask using binary cross-entropy, and the total loss is the sum

$\rightarrow$1

The paper attributes improved gradient flow to this deep supervision across the three subnetworks (Liu et al., 2023).

T-UNet uses Sigmoid-BCE plus Dice, Adam with initial learning rate $\rightarrow$2, and a decision threshold of $\rightarrow$3 at inference. The training description states that the model was trained on an RTX 3090 with batch size chosen to fit $\rightarrow$4 GB (Zhong et al., 2023).

The sequential 3D U-Net cascade uses binary soft Dice loss for all three modules, Nesterov-accelerated Adam with initial learning rate $\rightarrow$5, sagittal flips, and a patch sampling strategy of $\rightarrow$6 patches per subject with proportions $\rightarrow$7 background, $\rightarrow$8 normal brain, and $\rightarrow$9 tumour. Early stopping is applied at plateau or $\rightarrow$0 epochs maximum (Beers et al., 2017).

5. Application domains and reported performance

The triplet design has been instantiated across lesion segmentation, histology, neuro-oncology, and remote sensing, with each domain emphasizing a different reason for using three coordinated streams.

Paper	Domain	Reported results
(Sundaresan et al., 2021)	Brain tumour segmentation	Dice $\rightarrow$1 for ET and TC, Dice $\rightarrow$2 for WT on BraTS validation; mean Dice $\rightarrow$3, $\rightarrow$4, $\rightarrow$5 on ET, WT, TC for BraTS'20 unseen test
(Ahmed et al., 2024)	Nuclei instance segmentation on CryoNuSeg	Benchmark AJI $\rightarrow$6, PQ $\rightarrow$7; Triple U-Net AJI $\rightarrow$8, PQ $\rightarrow$9
(Liu et al., 2023)	Skin lesion segmentation on ISIC-2018	Dice $3\times 3$0, Jaccard $3\times 3$1, Accuracy $3\times 3$2; Params $3\times 3$3 M; FLOPs $3\times 3$4 G; FPS $3\times 3$5
(Zhong et al., 2023)	Change detection	Total parameters $3\times 3$6 M; FLOPs $3\times 3$7 G; best ablation setting $3\times 3$8, $3\times 3$9, $\rightarrow$0 on LEVIR-CD / WHU-CD / DSIFN-CD
(Beers et al., 2017)	Glioma tissue segmentation	Dice $\rightarrow$1 for WT, $\rightarrow$2 for ET, $\rightarrow$3 for TC

The histology model couples its three-branch network with watershed post-processing: Gaussian smoothing of the soft map, distance transform, local maxima as markers, and marker-based watershed on the inverted distance confined by the binarized mask. The description states that this substantially improves separation of touching nuclei (Ahmed et al., 2024). In the triplanar brain-tumour model, the independent TC branch is used in post-processing to reassign small ET islands, label voxels in $\rightarrow$4 as NCR/NET, and apply minor morphological clean-ups (Sundaresan et al., 2021).

For TM-UNet, the abstract reports experiments on ISIC17, ISIC18, CVC-300, CVC-ClinicDB, Kvasir-SEG, CVC-ColonDB, and Kvasir-Instrument, and states that the proposed model demonstrates superior segmentation performance while achieving a one-third reduction in parameters compared to the previous VM-UNet (Tang et al., 2024).

6. Conceptual boundaries and trajectory

A recurrent misconception is that Triplet-UNet denotes a single standardized architecture. The cited literature shows the opposite. In one usage, the triplet is a set of axial, sagittal, and coronal predictors whose outputs are averaged (Sundaresan et al., 2021). In another, it is a set of RGB, Hematoxylin, and segmentation branches fused by PDFA (Ahmed et al., 2024). In another, it is a three-stage refinement pipeline centered on ROIE and cross-stage skip reuse (Liu et al., 2023). In remote sensing, it is a triplet encoder over $\rightarrow$5, $\rightarrow$6, and $\rightarrow$7 combined through MBSSCA, CAM, and SAM (Zhong et al., 2023). In sequential 3D tumour segmentation, it is a biologically structured tree of WT, ET, and TC predictors (Beers et al., 2017). In TM-UNet, it is a Mamba-based system that uses residual VSS Blocks and Triplet SSM for spatial-channel fusion (Tang et al., 2024).

What unifies these systems is therefore not a fixed block inventory but a repeated design principle: three coordinated streams are used to separate complementary evidence and then recombine it through averaging, conditioning, progressive aggregation, attention, or cascade control. This suggests that Triplet-UNet is best understood as a structural motif for decomposing difficult dense-prediction problems rather than as a single architectural specification. The persistence of post-processing in several variants—tumour-core refinement in the triplanar ensemble and watershed in the nuclei model—also indicates that three-stream representation learning is often paired with explicit output regularization rather than treated as a fully sufficient end-to-end mechanism (Sundaresan et al., 2021, Ahmed et al., 2024).