SAEL: Semantic-Preserving Evidence Learning

• SAEL is a method that preserves semantically meaningful evidence, ensuring uncertainty maps reliably capture ambiguous regions in medical images.
• It integrates a semantic-smooth evidence generator with a fidelity-enhancing regularization term to mitigate the excessive suppression seen in standard evidential deep learning.
• Embedded within Evidential U-KAN, SAEL improves uncertainty-guided attention, boosting the model’s ability to handle blurred boundaries while incurring a modest segmentation accuracy trade-off.

SAEL, short for Semantic-Preserving Evidence Learning, is the evidence-learning component of Evidential U-KAN, introduced to improve the quality of evidence and uncertainty estimation in ambiguous, ill-defined, and boundary regions of medical images. Its central purpose is not merely to improve segmentation logits, but to preserve semantically meaningful ambiguity so that uncertainty maps remain faithful and can serve as active guidance signals in the paper’s progressive evidence uncertainty guided attention mechanism, denoted PEUA or EUGA. In this formulation, distorted uncertainty maps lead to distorted attention guidance; SAEL is therefore positioned as a corrective to a specific limitation of standard evidential deep learning (EDL), especially the tendency of KL regularization to suppress non-ground-truth evidence too aggressively and indiscriminately in boundary or class-confusing regions (Yang et al., 10 Oct 2025).

1. Conceptual role and motivation

SAEL is introduced against the background of trustworthy medical image segmentation, where the paper argues that many EDL-based methods neglect leveraging uncertainty maps rich in attention cues to refine ambiguous boundary segmentation. The paper’s criticism is directed at the standard EDL objective: although reducing evidence on incorrect classes is reasonable in unambiguous settings, it becomes harmful at blurred boundaries and class-confusing pixels, where some support for alternative classes is semantically meaningful and should translate into higher uncertainty, not be annihilated (Yang et al., 10 Oct 2025).

This yields the central contrast. A baseline EDL objective encourages evidence only for the labeled class and suppresses alternatives via KL regularization. SAEL instead preserves semantically meaningful ambiguity so that uncertainty remains informative in hard regions. The paper states that traditional regularization “indiscriminately suppress[es] the evidence values of incorrect classes,” which damages uncertainty assessment in ambiguous areas and, because uncertainty later drives PEUA/EUGA, also distorts the corresponding attention guidance. A plausible implication is that SAEL is not separable from the rest of Evidential U-KAN: it is the mechanism that makes uncertainty-guided refinement trustworthy enough to be useful.

2. Evidential formulation and failure mode of standard EDL

SAEL is built on the standard evidential segmentation formulation in which the network predicts evidence rather than direct softmax probabilities. For a $C$-class pixel $x$, the paper uses

$$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$

and

$$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$

Thus, larger total evidence implies larger $S$ and lower uncertainty $u_x$, while smaller total evidence implies higher uncertainty (Yang et al., 10 Oct 2025).

Within this formulation, the paper gives the evidential classification loss

$$\mathcal{L}_{CE}\left(p_x^C, y_x^C\right) = \prod_{c=1}^{C} \left(p_x^c\right)^{y_x^c}$$

and the evidential version

$$\mathcal{L}_{\mathrm{ice}} = \sum_{c=1}^{C} y_x^c \left(\log(S) - \log(\alpha_x^c)\right).$$

The accompanying notation is not fully cleanly specified in the source: $b_x^c$ appears in the loss notation but is not explicitly defined, and the indexing inside the Dirichlet normalization term is somewhat inconsistent. The paper nevertheless makes the intended role of the loss clear: it supervises evidential classification while SAEL changes how evidence is generated and how evidence and uncertainty are regularized.

The standard KL regularizer is written as

$$\begin{aligned} \mathcal{L}_{KL}(\boldsymbol{\alpha}_x^c) = &\log \left( \frac{\Gamma \left( \sum_{c=1}^{C} \tilde{\alpha}_{i,c} \right)} {\Gamma(C) \prod_{c=1}^{C} \Gamma(\tilde{\alpha}_{i,c})} \right) \ &+ \sum_{c=1}^{C} (\tilde{\alpha}_x^c - 1) \left[ \psi(\tilde{\alpha}_x^c) - \psi \left( \sum_{c=1}^{C} \tilde{\alpha}_x^c \right) \right] \end{aligned}$$

with

$x$0

The paper states that this adjustment ensures that the ground-truth evidence does not incorrectly become zero. The substantive critique, however, is not the formal validity of the KL term but its effect: globally pushing non-ground-truth evidence downward without regard to semantic ambiguity (Yang et al., 10 Oct 2025).

3. Constituent mechanisms of SAEL

SAEL consists of two explicit ingredients: a semantic-smooth evidence generator and a fidelity-enhancing regularization term. The evidence generator is

$x$1

where $x$2 is the network output before evidence conversion. The paper emphasizes that, as input increases, the gradient “gradually rises and eventually converges to 1” (Yang et al., 10 Oct 2025).

From the stated form, several properties follow directly. If $x$3, then $x$4 and $x$5, so evidence is nonnegative and zero on negative activations. If $x$6, evidence increases as $x$7, and

$x$8

This suggests that small positive activations generate evidence cautiously, whereas strong activations can still accumulate large evidence. In the paper’s terminology, this is why the generator is “semantic-smooth”: it avoids rapid evidence collapse for non-ground-truth classes, preserves semantically meaningful confusion at class boundaries, and thereby produces more faithful uncertainty.

The second mechanism is the fidelity-enhancing regularization term

$x$9

where $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$0 is the uncertainty at pixel $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$1 and $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$2 is the predicted probability of the ground-truth class. The paper says this term is minimized “when both the uncertainty and the predicted probability tend toward 1.” As noted in the source, that statement is mathematically imprecise over $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$3, but the intended role is explicit: when the model is not confident in the ground-truth class, optimization may proceed by maintaining or increasing uncertainty, rather than by indiscriminately suppressing competing evidence (Yang et al., 10 Oct 2025).

The final SAEL-related loss is

$$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$4

with

$$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$5

Accordingly, $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$6 is scheduled over training, whereas $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$7 is fixed. The paper does not describe SAEL as changing supervision labels; it changes evidence generation and modifies regularization behavior.

4. Integration into Evidential U-KAN

SAEL is integrated into both the training objective and the evidence generation stage of Evidential U-KAN. The paper’s pipeline is: an input image $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$8 passes through the U-KAN-based encoder-decoder; EUGA/PEUA uses the uncertainty map from the previous iteration to guide shallow features; the decoder produces logits; these logits are passed through the SAEL evidence generator; evidence is converted to Dirichlet parameters and uncertainty; the uncertainty map is fed to the next iteration of uncertainty-guided attention; and training uses the total loss above (Yang et al., 10 Oct 2025).

This positioning is important. SAEL affects training through $$\alpha_x^c = e_x^c + 1, \qquad S = \sum_{c=1}^{C} \alpha_x^c$$9 and $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$0, but it also affects inference, because the semantic-smooth evidence generator still determines how logits become evidence, and therefore changes the uncertainty map used in iterative prediction. The paper is explicit that SAEL is not merely a training trick.

Its interaction with PEUA/EUGA is the system-level rationale for the module. PEUA/EUGA relies on uncertainty maps as guidance; SAEL aims to make those maps more semantically faithful. A plausible implication is that SAEL’s primary value lies not only in uncertainty as an output for clinical interpretation, but also in uncertainty as an internal control signal for progressive feature refinement. The paper does not provide pseudocode or a dedicated architectural block diagram specifically for the semantic-smooth generator, and it appears to operate directly on the decoder output logits before evidence and Dirichlet conversion.

5. Empirical behavior, trade-offs, and reliability

The ablation on CVC-ClinicDB isolates EUGA and SAEL. Without either component, the model reports Dice $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$1, IoU $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$2, ASSD $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$3, and UEO $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$4. With SAEL alone, Dice becomes $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$5, IoU $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$6, ASSD $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$7, and UEO $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$8. With both EUGA and SAEL, the model reports Dice $$\hat{p}_x^c = \frac{\alpha_x^c}{S}, \qquad u_x = \frac{C}{S}.$$9, IoU $S$0, ASSD $S$1, and UEO $S$2 (Yang et al., 10 Oct 2025).

These results establish the paper’s main SAEL-specific trade-off. SAEL alone improves UEO from $S$3 to $S$4, but hurts segmentation accuracy and boundary quality when used without EUGA. When combined with EUGA, it gives the best reported uncertainty quality, $S$5, while maintaining strong segmentation. The paper explicitly states that SAEL causes a modest decline in segmentation accuracy but considerably augments the model’s capacity to discern uncertainty, improving trustworthiness.

The comparison with EDL-based uncertainty modeling on CVC-ClinicDB reinforces this interpretation. On the original test set, EDL + U-KAN reports $S$6, SAEL + U-KAN $S$7, EDL + Evi U-KAN $S$8, and SAEL + Evi U-KAN $S$9. On the noisy test set, the corresponding UEO values are $u_x$0, $u_x$1, $u_x$2, and $u_x$3. The same noisy comparison shows improved segmentation robustness for SAEL-based Evi U-KAN over EDL-based Evi U-KAN: Dice $u_x$4 versus $u_x$5, IoU $u_x$6 versus $u_x$7, ASSD $u_x$8 versus $u_x$9, and UEO $$\mathcal{L}_{CE}\left(p_x^C, y_x^C\right) = \prod_{c=1}^{C} \left(p_x^c\right)^{y_x^c}$$0 versus $$\mathcal{L}_{CE}\left(p_x^C, y_x^C\right) = \prod_{c=1}^{C} \left(p_x^c\right)^{y_x^c}$$1 (Yang et al., 10 Oct 2025).

The broader paper reports that, compared with U-KAN, the full Evidential U-KAN improves on average by Dice $$\mathcal{L}_{CE}\left(p_x^C, y_x^C\right) = \prod_{c=1}^{C} \left(p_x^c\right)^{y_x^c}$$2, IoU $$\mathcal{L}_{CE}\left(p_x^C, y_x^C\right) = \prod_{c=1}^{C} \left(p_x^c\right)^{y_x^c}$$3, and uncertainty evaluation metric $$\mathcal{L}_{CE}\left(p_x^C, y_x^C\right) = \prod_{c=1}^{C} \left(p_x^c\right)^{y_x^c}$$4. Those gains belong to the full method rather than to SAEL in isolation. The cautious interpretation given in the source is that SAEL is one of the key contributors to the uncertainty-quality side of these improvements, especially through UEO and robustness experiments. The paper also states that the full method has higher computational cost due to its iterative nature, but does not isolate SAEL’s overhead specifically.

6. Acronym ambiguity and adjacent usages

In recent arXiv usage, SAEL is not unique to medical image segmentation. The acronym also denotes “Leveraging LLMs with Adaptive Mixture-of-Experts for Smart Contract Vulnerability Detection,” a function-level vulnerability-detection framework that combines LLM predictions, LLM-generated explanations, prompt-tuned CodeT5 and T5, and an Adaptive Mixture-of-Experts architecture for reentrancy, timestamp dependency, integer overflow/underflow, and delegatecall detection (Yu et al., 30 Jul 2025).

A further source of confusion appears in SAEBench, where the accompanying note states that “SAEL should be read as sparse autoencoders for language-model interpretability,” in the context of a benchmark for sparse autoencoders rather than a distinct named method called SAEL (Karvonen et al., 12 Mar 2025). The visually similar acronym SEAL refers to the Santa Cruz Extreme AO Lab, a visible-wavelength extreme adaptive optics laboratory testbed for wavefront sensing and control on large segmented ground-based telescopes (Jensen-Clem et al., 2021).

For arXiv-reading audiences, this ambiguity matters because the segmentation meaning of SAEL is tightly tied to evidential uncertainty modeling, whereas the smart-contract meaning of SAEL concerns LLM-based vulnerability detection, and SEAL belongs to adaptive optics. In the medical-imaging literature represented here, SAEL refers specifically to the pair of mechanisms—semantic-smooth evidence generation and fidelity-enhancing uncertainty regularization—introduced to avoid evidence compression in traditional EDL at boundary regions and to support progressive uncertainty-guided attention (Yang et al., 10 Oct 2025).