Transparent Medical Attention (TMA)

• Transparent Medical Attention (TMA) is a family of approaches that expose and inspect model focus, intermediate reasoning, and workflow processes in clinical AI.
• It integrates intrinsic architectural mechanisms—such as spatial, channel, and scale attention—with operational transparency in medical imaging, dialogue, and pathology.
• TMA employs diverse strategies, from visual weight maps to auditable provenance workflows, to bolster trust and facilitate clinical review of AI outputs.

Transparent Medical Attention (TMA) is best understood as a heterogeneous research program rather than a single standardized method. Across the cited literature, the common objective is to make medically relevant focus, intermediate reasoning, or operational traces inspectable: in medical imaging through explicit spatial, channel, scale, region, or cross-modal attention; in medical dialogue through auditable emotional-state inference; and in clinical AI infrastructure through provenance, documentation, and interactive workflow control (Xie et al., 2023, Gu et al., 2020, Yang et al., 2024, Srinivasan et al., 12 Sep 2025, Imanov et al., 27 Jan 2026). In this sense, TMA denotes a family of approaches that seek to expose what a system is attending to, how that focus is constrained, and how its outputs can be reviewed in clinical or research settings.

1. Conceptual scope and terminological landscape

The underlying notion of attention in medical AI is relatively stable even when the term TMA is not. A general formulation given in the segmentation survey writes attention as

${\rm Attention} = f(g(x), x),$

where $g(x)$ is an attention signal and $f(g(x),x)$ denotes processing the input according to that signal (Xie et al., 2023). In the same survey, attention is described as a mechanism that adaptively assigns weights to different regions in an image and can “illustrate what the model has learned,” thereby providing interpretability insight into otherwise opaque neural networks (Xie et al., 2023). That framing supplies the conceptual core of TMA: not merely improving prediction, but exposing selective emphasis.

The literature, however, uses related language in several non-identical ways. In segmentation, CA-Net treats explainability as explicit awareness of important spatial positions, channels, and scales within a CNN, with attention weight maps serving as intrinsic visual evidence (Gu et al., 2020). In restoration, the Region Attention Transformer treats transparency as semantically constrained self-attention: tokens may attend only within explicit SAM-derived regions, making the support of attention visible and inspectable (Yang et al., 2024). In brain-tumor segmentation, TMA-TransBTS uses “TMA” to denote a 3D multi-scale attention design built from the 3D Multi-scale Self-attention Module and 3D Multi-scale Cross-attention Module, rather than a general transparency standard (Huang et al., 12 Apr 2025). A plausible implication is that “Transparent Medical Attention” functions more as an umbrella descriptor for transparency-oriented attention practices than as a single canonical architecture.

This terminological heterogeneity extends beyond neural attention. In digital pathology, TMA-Grid uses “TMA” in the established sense of Tissue MicroArray while emphasizing an interactive, inspectable, FAIR de-arraying workflow (Ge et al., 2024). In clinical NLP, TeMLM treats transparency as provenance, data transparency, modeling transparency, and governance rather than as attention-map inspection (Imanov et al., 27 Jan 2026). In medical dialogue, RECAP does not expose transformer heads, yet it makes patient-state interpretation explicit through staged emotional reasoning and per-emotion Likert signals (Srinivasan et al., 12 Sep 2025). TMA therefore spans architectural transparency, interactional transparency, and infrastructural transparency.

2. Intrinsic architectural transparency in medical imaging

A central line of TMA research builds transparency directly into image models. CA-Net is a canonical example. It augments a U-Net backbone with four spatial attention modules, four channel attention modules, and one scale attention module, aiming to make segmentation “aware of the most important spatial positions, channels and scales at the same time” (Gu et al., 2020). Its joint spatial attention combines a non-local block at the lowest resolution with dual-pathway attention gates at higher resolutions; channel attention recalibrates feature responses after encoder-decoder fusion; scale attention explicitly weights multiscale decoder outputs before prediction. The paper reports that CA-Net improved the average segmentation Dice score from 87.77% to 92.08% for skin lesion, 84.79% to 87.08% for the placenta, and 93.20% to 95.88% for the fetal brain compared with U-Net, while reducing model size to around 15 times smaller than DeepLabv3+ with close or better accuracy (Gu et al., 2020). In TMA terms, its importance lies not only in performance but in exposing spatial and scale weight maps as part of ordinary inference.

The survey of attention mechanisms in medical image segmentation situates such designs within a larger taxonomy of non-Transformer attention and Transformer attention (Xie et al., 2023). It distinguishes channel, spatial, and temporal attention among non-Transformer methods, then shows how these may be placed in the encoder, decoder, skip connections, or hybrid locations. This matters for TMA because the semantic meaning of an attention map depends heavily on placement. Decoder and skip-connection attentions are generally closer to final mask formation and therefore more legible as explanations of region selection, whereas encoder-side attentions are often more abstract (Xie et al., 2023). The same survey also formalizes self-attention in the usual scaled dot-product form,

$$g(x)= {\rm softmax}\left(\frac{QK^{T}}{\sqrt{d_k}}\right), \qquad f(g(x),x)=g(x)V,$$

making explicit that the interpretable object is the learned weighting structure itself (Xie et al., 2023).

TMA-TransBTS extends this architectural logic into 3D multi-modal MRI segmentation. The task uses four aligned volumetric modalities—T1, T1ce, T2, and FLAIR—and segments enhancing tumor, tumor core, and whole tumor (Huang et al., 12 Apr 2025). The model is an encoder-decoder network with four encoding stages and four decoding stages, operating on a 3D multi-modal MRI patch

$$\mathbf{X} \in \mathbb{R}^{H \times W \times D \times C},$$

with $C=4$ in the BraTS setting (Huang et al., 12 Apr 2025). Its 3D Multi-scale Self-attention Module uses a shared query

$\mathbf{Q}_{shared}=\mathbf{W}^Q\mathbf{I}$

together with scale-specific keys and values derived from multiple token granularities, and its 3D Multi-scale Cross-attention Module replaces ordinary skip fusion with cross-attentive encoder-decoder interaction (Huang et al., 12 Apr 2025). This design makes lesion scale itself an explicit attention object. A plausible implication is that TMA here consists not simply in visual heatmaps, but in structurally encoding the clinically salient fact that tumor appearance is multi-scale and volumetric.

3. Region-constrained attention, attribution, and output grounding

Another major TMA strand constrains attention through explicit image structure. The Region Attention Transformer for medical image restoration is the clearest formulation. It uses SAM to obtain masks

$$M = \operatorname{Postprocess}(\operatorname{SAM}(I^{LQ})),$$

postprocesses them into non-overlapping regions, resizes them to feature resolution, and constructs an additive attention mask

$$M_{ATT} = \lambda \sum_{i=1}^{L}\left(1 - M'_i {M'_i}^{\top}\right),$$

with $\lambda=-1000$ in experiments (Yang et al., 2024). Region-based self-attention is then computed as

$g(x)$0

The transparency claim is architectural rather than post hoc: cross-region interactions are masked out during the R-MSA step, so the admissible attention graph is semantically constrained and directly visualizable through the region masks (Yang et al., 2024). The same paper adds a focal region loss,

$g(x)$1

with $g(x)$2 and $g(x)$3, to emphasize difficult semantic regions during optimization (Yang et al., 2024). Quantitatively, it reports best results across PET synthesis, CT denoising, and pathological image super-resolution, including PSNR $g(x)$4 and SSIM $g(x)$5 for PET, PSNR $g(x)$6 and SSIM $g(x)$7 for CT, and PSNR $g(x)$8 and SSIM $g(x)$9 for pathological super-resolution (Yang et al., 2024).

Medical image captioning offers a related but distinct form of transparent grounding. The regional attention-enhanced Swin-BART system uses a Swin Transformer Base encoder producing

$f(g(x),x)$0

flattens this into 49 region vectors, computes scalar regional weights through a linear scoring function and softmax, and then reduces the visual sequence to 29 tokens before decoder cross-attention (Naz et al., 13 Nov 2025). The explicit regional attention mechanism is intended to amplify diagnostically salient regions before text generation, and the paper provides qualitative heatmaps for the regions driving each caption (Naz et al., 13 Nov 2025). On ROCO it reports ROUGE 0.603 and BERTScore 0.807, compared with ROUGE 0.356 and BERTScore 0.623 for a ResNet-CNN baseline and ROUGE 0.255 and BERTScore 0.645 for BLIP2-OPT, while remaining weaker on BLEU, CIDEr, and METEOR than the ResNet-CNN baseline (Naz et al., 13 Nov 2025). The resulting picture is that region-level transparency may improve semantic fidelity without guaranteeing uniformly superior lexical overlap.

Post-hoc inspection remains important in TMA even when the model is not intrinsically attention-based. A comparative XAI study across five medical datasets evaluates Integrated Gradients, Grad-CAM, and SHAP over ResNet50, DenseNet121, EfficientNetB3, and EfficientNetB0 classifiers (Wen et al., 20 Sep 2025). Its fidelity proxy is defined as

$f(g(x),x)$1

where $f(g(x),x)$2 is original confidence and $f(g(x),x)$3 is adversarial confidence under perturbation (Wen et al., 20 Sep 2025). The study repeatedly finds that explanation quality is tied to classifier quality: in brain tumor MRI and COVID-19 tasks, IG and SHAP often give more concentrated explanations than Grad-CAM, whereas in lower-performing lung cancer CT classification, explanations are less reliable overall (Wen et al., 20 Sep 2025). This supports a recurrent TMA theme: transparency methods cannot rescue weak underlying representations.

4. Transparent multimodal reasoning and patient-facing interaction

In dialogue systems, TMA shifts from visual grounding to explicit reasoning traces. RECAP—Reflect, Extract, Calibrate, Align, Produce—implements emotional alignment entirely at inference time and without retraining (Srinivasan et al., 12 Sep 2025). It stages response generation through situation abstraction, latent factor induction, candidate emotion extraction, Likert-based emotion likelihoods, and final emotionally aligned response production (Srinivasan et al., 12 Sep 2025). The core transparent object is a clinician-readable emotional profile in which each candidate emotion receives one of five ratings—very-unlikely, unlikely, neutral, likely, very-likely—with deterministic probability mapping

$f(g(x),x)$4

respectively (Srinivasan et al., 12 Sep 2025). This does not expose transformer attention weights, but it does expose what the system believes about the patient’s state, why it believes it, and how that belief shapes the response.

Empirically, RECAP improves emotional reasoning by 22–28% on 8B models and 10–13% on larger models over zero-shot baselines, with examples including EmoBench accuracy improving from 0.351 to 0.45 for Llama-3.1-8B and SECEU EQ score improving from 58.31 to 71.29 (Srinivasan et al., 12 Sep 2025). In blinded evaluation by two oncology research fellows, RECAP achieved a 95.7% win rate with 95% confidence interval $f(g(x),x)$5 across 186 decisive outcomes (Srinivasan et al., 12 Sep 2025). Its transparency is therefore tied to auditable intermediate artifacts rather than to mechanistic interpretability of model internals.

A broader multimodal version of this idea appears in Hulu-Med, which presents itself as a transparent generalist medical VLM spanning text, 2D images, 3D images, and video (Jiang et al., 9 Oct 2025). Its strongest transparency claim concerns release practice: open data curation, training algorithms, model parameters, and a medically aware token-reduction strategy that prunes redundant slices or frames in 3D and video (Jiang et al., 9 Oct 2025). The model was progressively trained on 16.7 million samples and is reported to require 4,000 to 40,000 GPU hours for 7B to 32B variants (Jiang et al., 9 Oct 2025). This is transparent pipeline design rather than explicit attribution, but it expands TMA from local attention maps to large-scale reproducible multimodal system construction.

Transparent medical attention also appears in direct sensing of human behavior. A multimodal learning-analytics study of the ABCDE procedure in nursing education links gaze, hand movement, proximity, and video-validated procedural annotations through behaviorgrams (Heilala et al., 2023). Gaze is summarized by Shannon joint entropy,

$f(g(x),x)$6

with lower entropy interpreted as more coherent, focused visual attention (Heilala et al., 2023). The study identifies four primary phases of the procedure and shows that focused gaze can occur either with strong hand activity or with minimal hand movement, distinguishing observational from manipulative subphases (Heilala et al., 2023). Here TMA is neither neural nor post hoc; it is a transparent sensing-and-visualization framework for procedural attention in medical education.

5. Provenance, workflow, and infrastructural transparency

A distinct but increasingly important form of TMA concerns workflow-level transparency. TeMLM formalizes this for clinical NLP through a machine-checkable release bundle composed of TeMLM-Card, TeMLM-Datasheet, and TeMLM-Provenance, organized around provenance, data transparency, modeling transparency, and governance (Imanov et al., 27 Jan 2026). Documentation completeness is defined as

$f(g(x),x)$7

missingness for structured field $f(g(x),x)$8 as

$f(g(x),x)$9

and drift sensitivity through Population Stability Index

$$g(x)= {\rm softmax}\left(\frac{QK^{T}}{\sqrt{d_k}}\right), \qquad f(g(x),x)=g(x)V,$$0

The framework insists that data entities, code entities, and model entities be versioned and linked by cryptographic hashes, and that even when raw clinical text cannot be shared, a minimal provenance bundle can still expose event logs, hashes, schemas, and aggregate statistics (Imanov et al., 27 Jan 2026). In TMA terms, the focus shifts from “what region was attended to” toward “what pipeline produced the model, under what governance constraints, and with what auditable lineage.”

Digital pathology offers a complementary operational example. TMA-Grid is an in-browser, zero-footprint web application for Tissue MicroArray de-arraying that combines a U-Net-like convolutional neural network for tissue segmentation with an interactive grid estimation algorithm (Ge et al., 2024). It generates a 512×512 representation for segmentation, thresholds and postprocesses the probability output, estimates rows and columns from core centroids, and then allows manual corrections of segmentation, rotation, radius, row/column assignments, search radius, and stopping distance, with real-time grid reconstruction after each edit (Ge et al., 2024). On a held-out test set of 12 TMA whole-slide images, its segmentation model achieved loss 0.224, AUC 0.981 in Table 1, accuracy 0.919, precision 0.818, and recall 0.947 (Ge et al., 2024). The application’s transparency lies in the fact that the algorithmic pipeline is not hidden behind a black box: users can inspect overlays, adjust hyperparameters, and export both core images and structured metadata (Ge et al., 2024).

Telehealth Chain extends infrastructural transparency into telemedicine. Built on Hyperledger Fabric, it proposes modular smart contracts for access control, communication, data storage, consent management, payments, and analytics (Mahdi et al., 2023). The goal is controlled transparency among legitimate stakeholders rather than public visibility: access requests, consent grants or revocations, communication events, and payment transactions become immutable, auditable ledger entries (Mahdi et al., 2023). This model targets security, privacy, auditability, integrity, and non-repudiation in remote care workflows. It therefore contributes to TMA at the governance layer by making attention events in the broad sense—who accessed a record, who communicated, who paid, and under what authorization—verifiable after the fact.

6. Limitations, controversies, and adjacent clinical meanings

A recurrent misconception is that attention is automatically explanation. The surveyed literature does not support that conclusion. The segmentation survey states that attention can provide interpretability insight, but it also emphasizes future challenges in robustness, task specificity, and standard evaluation, and notes that attention results may be partially or completely incorrect (Xie et al., 2023). CA-Net provides intrinsic spatial, channel, and scale weight maps, yet its own claims are best read as improved transparency and introspection rather than definitive causal explanation (Gu et al., 2020). RAT explicitly frames its transparency as architectural and visual, not as a formal interpretability theory or user study (Yang et al., 2024). The medical image captioning system provides attention heatmaps but acknowledges the absence of rigorous faithfulness validation and region-level supervision (Naz et al., 13 Nov 2025). RECAP, similarly, improves empathy through explicit staged reasoning without exposing internal neural attention heads (Srinivasan et al., 12 Sep 2025). TMA therefore contains a persistent controversy between plausible, inspectable focus representations and formally validated explanation.

Another limitation is that transparency mechanisms are often domain- and task-specific. The restoration results of RAT depend on SAM-derived semantic partitions, so segmentation errors in SAM can misguide the entire attention scope (Yang et al., 2024). TMA-Grid’s gridding assumptions can fail under severe nonlinear distortion or unconventional TMA layouts, and its validation remains modest and primarily qualitative beyond segmentation metrics (Ge et al., 2024). TeMLM’s demonstration dataset, Technetium-I, is synthetic and is explicitly presented as suitable for tooling and process validation but insufficient for deployment without governed real-world validation (Imanov et al., 27 Jan 2026). RECAP imposes approximately 3–5× more token processing than baseline prompting and is vulnerable to error propagation across its stages (Srinivasan et al., 12 Sep 2025). Transparent systems, in other words, often trade simplicity of deployment for inspectability and governance.

The literature also shows that “transparency” may refer to clinical communication rather than model introspection. An adjacent example is the study of medical, cloth, and transparent face masks on speech signals (Corey et al., 2020). It finds that transparent masks preserve visual information but have poor acoustic performance compared with standard medical masks, with transparent window masks attenuating roughly 8 dB for the human talker and 10–14 dB for the loudspeaker, while lapel microphones are minimally affected by most masks (Corey et al., 2020). The resulting tradeoff—visual access versus degraded high-frequency speech cues—illustrates that transparent medical communication is itself a technical problem. A plausible implication is that TMA, understood broadly, includes not only transparent model focus but also the transparent transmission of clinically relevant signals between humans.

The likely future trajectory of TMA is therefore plural rather than singular. One direction strengthens intrinsic architectural transparency through better region, boundary, scale, and cross-modal constraints. Another extends auditable reasoning traces in dialogue and multimodal systems. A third builds end-to-end transparency through provenance, documentation, interactive correction, and verifiable transaction logs. The common requirement across these variants is not the existence of an “attention map” as such, but the availability of technically meaningful, clinician-reviewable evidence about what a medical system emphasized, how that emphasis was formed, and how it affected downstream action (Xie et al., 2023, Srinivasan et al., 12 Sep 2025, Imanov et al., 27 Jan 2026).