HANS-Net: Segmentation & Anomaly Detection

• HANS-Net is a dual-purpose framework that enhances CT liver and tumor segmentation using hyperbolic convolutions, adaptive temporal attention, and wavelet-inspired decomposition.
• It embeds heterogeneous anomaly detectors through a bagged approach, exposing Clever Hans behavior by combining density, reconstruction, and boundary-based analyses.
• The architecture delivers robust performance with improved Dice scores and generalization on LiTS and 3D-IRCADb-01, while offering enhanced boundary refinement and uncertainty estimation.

Searching arXiv for papers on HANS-Net and closely related "Clever Hans" methods. HANS-Net designates two distinct constructions in recent arXiv literature. In medical image analysis, it denotes Hyperbolic-convolutions Adaptive-temporal-attention with Neural-representation and Synaptic-plasticity Network, a hybrid framework for liver and tumor segmentation in abdominal CT that integrates wavelet-inspired decomposition, hyperbolic convolutions, adaptive temporal attention, synaptic plasticity, an implicit neural representation branch, and uncertainty-aware Monte Carlo dropout (Abian et al., 15 Jul 2025). In anomaly detection, the same name is used for a combined, explanation-enabled perspective that exposes and partially mitigates Clever Hans behavior by embedding heterogeneous detectors into a shared abstraction and averaging their scores (Kauffmann et al., 2020). The two usages are conceptually unrelated except for the shared concern with robustness and generalization.

1. Nomenclature and bibliographic scope

In the segmentation literature, HANS-Net is a named architecture introduced for accurate and generalizable liver and tumor segmentation in CT imaging (Abian et al., 15 Jul 2025). Its acronym expands to a system that combines hyperbolic convolutions, adaptive temporal attention, neural representation, and synaptic plasticity, with additional modules for multi-scale texture learning and uncertainty estimation. The design goal is explicitly anatomical: hierarchical structure, fine tumor boundaries, inter-slice continuity, and calibrated confidence are handled within a single end-to-end system.

In the anomaly-detection literature, HANS-Net is not a single detector architecture in the conventional sense. Rather, it names a framework that embeds density-based, reconstruction-based, and boundary-based anomaly detectors into a common feature extraction / distance computation / pooling view, explains them with LRP-style methods, and mitigates their individual structural weaknesses by bagging (Kauffmann et al., 2020). The purpose there is to study when detectors are right for the wrong reasons.

Several related papers on shortcut learning and the Clever Hans effect explicitly discuss the phenomenon without introducing a model named HANS-Net. This is stated for speech-based Alzheimer’s disease detection on the Pitt corpus, for preemptive pruning of hidden Clever Hans strategies via EGEM, and for unsupervised learning audits using LRP, BiLRP, and virtual frequency layers (Liu et al., 2024, Linhardt et al., 2023, Kauffmann et al., 2024). A plausible implication is that the string “HANS-Net” can produce bibliographic ambiguity: in one branch of the literature it is a proper architecture name, while in another it is associated with a wider vocabulary of shortcut-learning analysis.

2. Segmentation problem formulation and design rationale

The segmentation HANS-Net is motivated by three recurring deficiencies in prior liver/tumor segmentation methods: difficulty modeling complex anatomical hierarchy, poor capture of fine tumor boundaries and inter-slice continuity, and lack of confidence/uncertainty estimation (Abian et al., 15 Jul 2025). The underlying claim is that many existing systems still rely on Euclidean convolutions and “flattened” representations, which limits their ability to represent nested liver/tumor anatomy, irregular and thin tumor boundaries, continuity across adjacent CT slices, robust generalization to unseen datasets, and uncertainty in difficult regions.

The architecture is therefore organized around six complementary modules: wavelet-inspired decomposition, hyperbolic convolutions, adaptive temporal attention, synaptic plasticity, implicit neural representation, and Monte Carlo dropout uncertainty estimation (Abian et al., 15 Jul 2025). Each module contributes a different inductive bias. Wavelet-style decomposition targets multi-scale frequency structure; hyperbolic geometry targets hierarchical organization; temporal attention targets inter-slice consistency; synaptic plasticity targets adaptive feature strengthening; the implicit neural representation branch targets continuous boundary refinement; and Monte Carlo dropout targets predictive uncertainty.

This modular composition is significant because it does not frame segmentation as a single representational problem. Instead, the network treats organ-scale shape, lesion-scale texture, slice-to-slice coherence, and confidence calibration as distinct but coupled subproblems. The paper’s interpretation is that better segmentation follows from aligning the geometry of the representation space, the frequency content of the input, and the topology of the output contours (Abian et al., 15 Jul 2025).

3. Architectural composition and mathematical formulation

The paper describes the pipeline as: input CT slice/feature map, then Wavelet Decomposition Module (WDM), then Hyperbolic Convolution blocks, then Adaptive Temporal Attention (ATA), then Synaptic Plasticity Module (SPM), then an Implicit Neural Representation (INR) branch, and finally an Uncertainty Estimator via Monte Carlo dropout (Abian et al., 15 Jul 2025). The modules are presented as complementary rather than interchangeable.

The Wavelet Decomposition Module receives an input tensor

$$\mathbf{x} \in \mathbb{R}^{B \times C \times H \times W},$$

and extracts low-pass and directional high-pass components via convolutional filters:

$$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$

These are fused by a $1\times1$ convolution. In the paper’s interpretation, the low-frequency branch preserves global organ context, while the high-frequency branches emphasize edges and texture changes, which are especially important for tumors (Abian et al., 15 Jul 2025).

The Hyperbolic Convolution module first applies a conventional convolution and then maps features into hyperbolic space. With transformed feature tensor $\mathbf{F}'$ and learnable matrix $\mathbf{T}$, the mapping is

$\mathbf{Z} = \exp_\kappa(\mathbf{F}' \mathbf{T}),$

with

$$\exp_\kappa(\mathbf{v}) = \frac{ \tanh\!\big( \sqrt{-\kappa} \, \| \mathbf{v} \| \big) }{ \sqrt{-\kappa} \, \| \mathbf{v} \| + \epsilon } \mathbf{v}.$$

The hyperbolic features are then reshaped back to image-grid form. The stated rationale is that hyperbolic geometry can represent hierarchical, nested, tree-like structure more compactly than Euclidean feature spaces (Abian et al., 15 Jul 2025).

The Adaptive Temporal Attention block is intended to model inter-slice dependencies without a heavy 3D transformer. It uses a query-key-value attention mechanism over spatial features and modulates the attention output with a learned temporal weight vector. The effect claimed in the paper is smoother transitions across slices and fewer isolated false positives or false negatives (Abian et al., 15 Jul 2025).

The Synaptic Plasticity Module is biologically inspired. For each sample, the feature tensor is globally averaged to obtain channel activity, then converted into a channel correlation matrix, and the synaptic trace is updated by

$$\boldsymbol{\eta}' \leftarrow \boldsymbol{\eta} + \alpha \, \text{diag}(\mathbf{C}) - \boldsymbol{\eta}.$$

The output is then modulated as

$$\mathbf{F}_\text{out} = \mathbf{W}_\text{conv} * \big[ (\mathbf{X} \odot (\boldsymbol{\omega} \odot (1 + \boldsymbol{\eta}))) \mathbf{T} \big].$$

The intended effect is dynamic strengthening of channels that are consistently useful, thereby improving adaptive feature emphasis in tumor-relevant regions (Abian et al., 15 Jul 2025).

The Implicit Neural Representation branch converts segmentation from fixed-grid prediction into coordinate-conditioned prediction. For normalized coordinates $\mathbf{p}\in[-1,1]^2$, the positional encoding is

$$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$0

A sampled local feature vector $$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$1 is concatenated with coordinates and positional encoding, and an MLP

$$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$2

predicts logits:

$$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$3

This gives continuous-resolution predictions and is intended to improve boundary smoothness and anatomical plausibility (Abian et al., 15 Jul 2025).

The Monte Carlo dropout branch supplies uncertainty estimates. At test time, dropout is applied across $$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$4 stochastic forward passes:

$$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$5

The mean prediction and predictive variance are

$$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$6

The resulting uncertainty maps are intended to highlight ambiguous regions, particularly heterogeneous tumors and boundaries (Abian et al., 15 Jul 2025).

4. Training protocol, preprocessing, and evaluation setting

The segmentation HANS-Net is trained end-to-end, and the paper does not present a custom loss beyond standard segmentation training (Abian et al., 15 Jul 2025). During training, the components are chained in the order specified by the architecture. During inference, the uncertainty branch requires repeated dropout-enabled forward passes, after which output averaging and variance estimation are performed.

The evaluation setting uses LiTS for training/validation/testing and 3D-IRCADb-01 for cross-dataset validation (Abian et al., 15 Jul 2025). Preprocessing includes CT windowing to $$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$7, rescaling to $$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$8, resampling, and resizing slices to $$\mathbf{F} = \text{Concat}\Big( \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{low}), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^h), \text{Conv2D}(\mathbf{x}, \mathbf{W}_\text{high}^v) \Big).$$9. These choices are important because the architecture combines image-grid feature extraction with coordinate-based continuous decoding; the preprocessing pipeline therefore constrains both the scale of texture features and the normalization of spatial coordinates.

The paper evaluates four standard segmentation metrics: Dice, IoU, ASSD, and VOE (Abian et al., 15 Jul 2025). Dice and IoU quantify overlap, ASSD quantifies average boundary distance in millimeters, and VOE quantifies volume overlap error. The paper treats their joint use as necessary because the architecture is designed not only to improve regional overlap but also to refine contours and stabilize predictions across slices.

5. Reported performance, cross-dataset generalization, and ablation

On LiTS, the paper reports liver segmentation scores of Dice 96.67, IoU 93.79, ASSD 0.54 mm, and VOE 6.21, and tumor segmentation scores of Dice 89.84, IoU 82.39, ASSD 0.89 mm, and VOE 17.61. The corresponding mean values are Dice 93.26, IoU 88.09, ASSD 0.72 mm, and VOE 11.91 (Abian et al., 15 Jul 2025). In the abstract, the same overall LiTS result is summarized as a mean Dice score of 93.26%, an IoU of 88.09%, an average symmetric surface distance of 0.72 mm, and a volume overlap error of 11.91%.

For 3D-IRCADb-01, the paper reports liver scores of Dice 94.82, IoU 90.85, ASSD 0.97 mm, and VOE 9.15, and tumor scores of Dice 80.08, IoU 69.74, ASSD 2.08 mm, and VOE 30.26. The mean values are Dice 87.45, IoU 80.30, ASSD 1.53 mm, and VOE 19.71; the abstract reports the mean ASSD more precisely as 1.525 mm (Abian et al., 15 Jul 2025). The paper interprets this as evidence of strong generalization across different datasets, while also noting that tumor performance drops relative to LiTS due to dataset variability and tumor heterogeneity.

Relative to prior methods, the paper states that on LiTS the model outperforms U-Net, U-Net++, Attention U-Net, H-DenseUNet, and TransUNet on mean Dice/IoU and mean ASSD/VOE (Abian et al., 15 Jul 2025). It also notes that TransUNet had a slightly lower liver ASSD (0.51 mm) and slightly lower tumor VOE in some comparisons, but HANS-Net achieved the best overall overlap metrics and best mean performance. On 3D-IRCADb-01, HANS-Net is reported to achieve the best mean Dice, mean IoU, and lowest mean ASSD/VOE among the compared methods (Abian et al., 15 Jul 2025).

The ablation sequence is especially informative. The paper reports an A1 baseline with Dice 87.55, then progressive gains: A2 with + Hyperbolic Convolution reaches 89.60, A3 with + Wavelet Decomposition reaches 90.79, A4 with + Adaptive Temporal Attention reaches 91.89, A5 with + Synaptic Plasticity reaches 92.60, A6 with + INR reaches 93.18, and A7 with + Uncertainty Estimator reaches 93.86 (Abian et al., 15 Jul 2025). The paper identifies Hyperbolic Convolution as the largest single boost, interprets Wavelet Decomposition as improving multiscale detail sensitivity, Adaptive Temporal Attention as improving inter-slice consistency and boundary precision, INR as improving continuous boundary fidelity, and the Uncertainty Estimator as improving reliability, confidence estimation, and final measured performance.

6. Relation to the Clever Hans literature

The phrase Clever Hans predictor refers to a model that is statistically effective on a benchmark but conceptually wrong, brittle, or non-generalizable because it exploits shortcuts, artifacts, or loopholes rather than task-relevant structure (Lapuschkin et al., 2019). This line of work emphasizes that conventional metrics such as accuracy, mAP, reward, or ROC do not reveal how a model succeeds. Methods such as Layer-wise Relevance Propagation (LRP) and Spectral Relevance Analysis (SpRAy) were developed precisely to inspect model evidence and cluster explanation patterns at scale (Lapuschkin et al., 2019, Anders et al., 2019).

The anomaly-detection usage of HANS-Net belongs directly to this tradition. That work argues that anomaly detectors can be right for the wrong reasons, proposes a unified explanation framework across KDE, autoencoder, and deep one-class models, and defines a bagged score

$1\times1$0

after score standardization on training examples (Kauffmann et al., 2020). The aim is not to define a new monolithic backbone but to expose and partially cancel model-specific structural weaknesses.

Subsequent papers extend the same intellectual program without using the name HANS-Net. ClArC addresses dataset-scale discovery and removal of class artifacts through attribution analysis and concept-direction suppression (Anders et al., 2019). EGEM and PCA-EGEM attempt to preemptively prune hidden shortcut strategies even when explanations on clean validation data appear acceptable (Linhardt et al., 2023). In speech-based Alzheimer’s disease detection, near-perfect accuracy from silence-only segments on the Pitt corpus is presented as a textbook Clever Hans effect driven by background noise and silence-related cues, and the paper explicitly states that it does not define a named system called HANS-Net (Liu et al., 2024). In unsupervised learning, CH effects are shown to be widespread in CLIP, SimCLR, Barlow Twins, PubMedCLIP, D2Neighbors, and PatchCore, again without introducing a model called HANS-Net (Kauffmann et al., 2024).

This broader context clarifies the position of the CT-segmentation HANS-Net. It is a domain-specific segmentation architecture, not an explanation-based shortcut-auditing method. A plausible implication is that the shared name links two otherwise separate research trajectories: one concerned with anatomical representation and clinically reliable segmentation, the other with diagnosing and mitigating models that are accurate for the wrong reasons.