Hierarchical Tissue-aware Magnification

• HTM is a hierarchical magnification framework that integrates coarse-to-fine processing with tissue-aware localization for enhanced imaging analysis.
• It employs dual-stage mechanisms—initial broad tissue segmentation followed by focused magnification—to adaptively allocate resolution based on tissue morphology and motion.
• Empirical evaluations demonstrate improved metrics (e.g., SSIM, PSNR) and diagnostic accuracy, though its current computational profile limits real-time performance.

Searching arXiv for the specified papers to ground the article in current literature. Hierarchical Tissue-aware Magnification (HTM) denotes a class of coarse-to-fine magnification mechanisms that couple tissue- or structure-aware localization with selective high-resolution emphasis while preserving surrounding context. In current usage across recent arXiv literature, the term appears in several technically distinct but conceptually related forms: as a per-pixel spatial-conditioning module for endoscopic vascular motion magnification in EndoControlMag, as a question-guided tissue-aware retrieval hierarchy for whole-slide image reasoning in HistoSelect, as a differentiable multi-level zoom operator for multiple instance learning in ZoomMIL, and as focus+context multi-scale lensing in Scope2Screen (Wanga et al., 21 Jul 2025, Huang et al., 28 Feb 2026, Thandiackal et al., 2022, Jessup et al., 2021). Across these formulations, the common objective is not uniform enlargement, but structured allocation of magnification or representational bandwidth to regions whose morphology, motion, or semantic relevance justifies finer analysis.

1. Definitional scope and recurring structure

The most explicit formulation of HTM is given in EndoControlMag, where it is described as the core spatial-conditioning module responsible for generating a two-stage, biomechanically plausible magnification mask that tightly follows the vessel “core” through arbitrary camera and tissue motions and smoothly attenuates amplification into the surrounding tissue (Wanga et al., 21 Jul 2025). In HistoSelect, the analogous hierarchy is expressed as a “coarse-to-fine magnification” pipeline in which a group sampler performs a low-magnification scan to identify relevant tissue types and a patch selector performs a high-magnification zoom-in within those tissues (Huang et al., 28 Feb 2026). ZoomMIL formulates the same diagnostic intuition as multi-level zooming over a whole-slide pyramid, and Scope2Screen realizes it as interactive tissue-aware “lensing” spanning single-cell and tissue-neighborhood scales (Thandiackal et al., 2022, Jessup et al., 2021).

Setting	HTM formulation	Primary unit
Endoscopic video	Dual-mask spatial magnification with tracked vessel core and softened outer ring	Pixel mask
Pathology QA	Group sampler followed by patch selector	Patch/token
WSI MIL	Differentiable Top-K zooming across pyramid levels	Patch/region
Interactive visualization	Nested focus+context lenses with dynamic scale switching	Lens/ROI

Taken together, these works suggest a recurring architectural pattern: a coarse stage identifies a biologically or semantically relevant support, a fine stage increases resolution or weighting within that support, and a transition mechanism preserves continuity between focus and context. A plausible implication is that HTM is best understood as a design principle rather than a single algorithmic primitive.

2. EndoControlMag: per-pixel HTM in endoscopic vascular motion magnification

Within EndoControlMag, each incoming frame $I_t$ is paired with a periodically reset reference $I_r^k$ through the Periodic Reference Resetting (PRR) scheme, and with a positionally encoded magnification factor $\alpha$ (Wanga et al., 21 Jul 2025). HTM provides the per-pixel mask $M_t(\mathbf{x})$, which encodes a spatially varying weight in $[0,1]$ and is element-wise multiplied by the embedded $\alpha$ before being fed into the underlying Lagrangian magnifier $f_{\mathrm{VMM}}$.

The HTM pipeline proceeds in a fixed sequence. First, the inner mask $M_t^{\mathrm{in}}$ around the vessel core is updated via a pretrained Visual Object Tracking model. Second, HTM computes an adaptive dilation radius

$$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$

with $\gamma = 1/15$, and applies morphological dilation to obtain a broader region

$I_r^k$0

The outer ring is then defined as

$I_r^k$1

Third, HTM computes a softening weight $I_r^k$2 in the outer ring using one of two modes. Finally, it assembles the unified mask

$I_r^k$3

This mask is then fused with a positional embedding of $I_r^k$4 and passed to the VMM model to produce the magnified frame.

The dual-region construction is central. The inner mask $I_r^k$5 corresponds to the vessel core and is always magnified with full strength, whereas the outer mask $I_r^k$6 defines surrounding tissue whose magnification is attenuated rather than abruptly cut off. The paper explicitly states that this adaptive dilation ensures the transition width scales with vessel size rather than a one-size-fits-all radius (Wanga et al., 21 Jul 2025).

3. Tracking, softening modes, and biomechanical plausibility

HTM maintains vessel-core alignment by using MFT, “Long-term Tracking of Every Pixel,” as $I_r^k$7 (Wanga et al., 21 Jul 2025). Dense boundary or keypoints from $I_r^k$8 are tracked into $I_r^k$9, producing

$\alpha$0

This recursive propagation is reported to provide view-invariant alignment under camera pans and zooms, occlusion resilience because individual pixel-point trackers re-appear when the vessel re-enters view, and boundary precision by avoiding the “jumps” or mis-registration associated with static masks.

The outer-ring softening has two modes. In motion-based softening, when optical-flow estimation is reliable, optical flow is computed with RAFT:

$\alpha$1

and the weight is assigned proportionally to normalized displacement magnitude:

$\alpha$2

The effective local magnification factor is then $\alpha$3. In distance-based softening, intended for adverse conditions such as smoke, specular highlights, and flow breakdowns, each outer-ring pixel is weighted by an exponential decay:

$\alpha$4

with $\alpha$5, so that $\alpha$6 (Wanga et al., 21 Jul 2025).

The paper characterizes the two modes as complementary rather than competitive. Motion-based softening excels with complex tissue deformations, whereas distance-based softening provides stability during unreliable optical flow conditions. This dual-mode design is presented as the mechanism by which HTM remains robust across occlusions, instrument disturbance, view changes, and vessel deformations (Wanga et al., 21 Jul 2025).

4. Parameterization, empirical results, and computational profile

The reported HTM-related parameter settings in EndoControlMag are compact and task-specific. PRR uses clip length $\alpha$7 frames, chosen via ablation to balance drift against temporal coherence, with approximately $\alpha$8 windows capturing sub-cardiac cycles. The dilation ratio is $\alpha$9, and the decay constant is $M_t(\mathbf{x})$0; RAFT is used for optical flow and MFT for long-term mask tracking (Wanga et al., 21 Jul 2025).

Quantitative results are reported for both easy and hard settings. On the Easy Set of 4 static clips, FlowMag achieves $M_t(\mathbf{x})$1, whereas EndoControlMag (dist) reaches $M_t(\mathbf{x})$2; PSNR at $M_t(\mathbf{x})$3 improves from $M_t(\mathbf{x})$4 to $M_t(\mathbf{x})$5 for distance-based softening and $M_t(\mathbf{x})$6 for motion-based softening; motion-error $M_t(\mathbf{x})$7 at $M_t(\mathbf{x})$8 decreases from $M_t(\mathbf{x})$9 to $[0,1]$0 for distance mode and $[0,1]$1 for motion mode; magnification-error $[0,1]$2 at $[0,1]$3 decreases from $[0,1]$4 to $[0,1]$5 for distance mode and $[0,1]$6 for motion mode (Wanga et al., 21 Jul 2025).

On the Hard Set of 20 challenging clips, the reported gains over FlowMag include averaged SSIM improvements of $[0,1]$7–$[0,1]$8 across occlusion, view-change, deformation, and tool-disturbance scenarios, together with PSNR improvements of $[0,1]$9–$\alpha$0 (Wanga et al., 21 Jul 2025). Surgeon grading on a 10-item rubric further differentiates the methods: on easy clips, FlowMag scores $\alpha$1 versus EndoControlMag at approximately $\alpha$2; on hard clips, FlowMag scores $\alpha$3 versus EndoControlMag at approximately $\alpha$4.

The computational profile is correspondingly explicit. HTM requires one MFT forward pass, one RAFT pass in motion mode, and mask-morphology plus softening-weight computations per frame. On an NVIDIA RTX A6000, total per-frame time is approximately $\alpha$5. The paper states that this is insufficient for $\alpha$6–$\alpha$7 live feedback, but acceptable for offline review or guided procedures with intermittent feedback; it further notes model distillation, lighter flow or tracking backbones, and hardware-specific optimizations as possible routes toward real-time operation (Wanga et al., 21 Jul 2025).

5. HTM as coarse-to-fine retrieval and differentiable zooming in digital pathology

In computational pathology, HTM appears not as a per-pixel magnification mask but as hierarchical control over which tissue regions receive high-resolution attention. HistoSelect organizes this into three stages: tissue segmentation, hierarchical selection, and LLM answering (Huang et al., 28 Feb 2026). Patches are first embedded with CONCH, pathologist-defined tissue prompts are embedded with the same CLIP-style text encoder, and each patch receives a tissue label

$\alpha$8

which partitions the slide into tissue groups. The group sampler then computes

$\alpha$9

where $f_{\mathrm{VMM}}$0 is the question embedding. The patch selector predicts

$f_{\mathrm{VMM}}$1

and selects the top

$f_{\mathrm{VMM}}$2

patches within each tissue group. The selected embeddings are concatenated with question tokens and fed to a Qwen2.5-7B decoder. The paper explicitly maps these stages to magnification: the group sampler is the low-magnification scan, and the patch selector is the high-magnification zoom-in (Huang et al., 28 Feb 2026).

This formulation is quantitatively tied to efficiency. Peak performance occurs at approximately $f_{\mathrm{VMM}}$3 patches, roughly $f_{\mathrm{VMM}}$4 of all patches, yielding $f_{\mathrm{VMM}}$5 savings in visual tokens versus uniform sampling. Averaged over three closed-ended VQA benchmarks, Quilt-LLaVA and SlideChat obtain $f_{\mathrm{VMM}}$6 and $f_{\mathrm{VMM}}$7, whereas HistoSelect obtains $f_{\mathrm{VMM}}$8. On WSI-Bench open-ended report generation, BLEU-4 and ROUGE-L improve from $f_{\mathrm{VMM}}$9 to $M_t^{\mathrm{in}}$0, and WSI-Precision/Recall for diagnosis improve from $M_t^{\mathrm{in}}$1 to $M_t^{\mathrm{in}}$2. Diagnosis-task ablations report $M_t^{\mathrm{in}}$3 for random sampling, $M_t^{\mathrm{in}}$4 for DivPrune, $M_t^{\mathrm{in}}$5 for simple cosine similarity, and $M_t^{\mathrm{in}}$6 for the full HistoSelect model; removing the group sampler or patch selector reduces diagnosis accuracy to $M_t^{\mathrm{in}}$7 and $M_t^{\mathrm{in}}$8, respectively (Huang et al., 28 Feb 2026).

ZoomMIL addresses a related problem in multiple instance learning on whole-slide images by learning multi-level zooming end-to-end over a magnification pyramid (Thandiackal et al., 2022). At each magnification $M_t^{\mathrm{in}}$9, non-overlapping $$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$0 patches are embedded with a truncated ResNet-50, passed through dual gated attention modules, and used both for pooling into a slide-level representation $$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$1 and for selecting $$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$2 regions of interest to zoom further. The discrete selection matrix $$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$3 is defined by a linear program over the convex hull of valid sorted one-hot picks and made differentiable via a perturbed-optimizer construction:

$$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$4

Selected low-resolution regions are then expanded through the pyramid so that only embeddings corresponding to the chosen regions are retained at the next magnification. The final representation is

$$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$5

followed by classification with a two-layer MLP. The method is reported to outperform state-of-the-art MIL methods on two large datasets while reducing FLOPs and processing time by up to $$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$6 (Thandiackal et al., 2022).

A plausible implication of HistoSelect and ZoomMIL is that HTM can be implemented not only as an image-space magnification field, but also as a learned allocation policy over tokens, patches, or instances, provided that the hierarchy preserves tissue semantics across scales.

6. Focus+context lensing, validation, and limitations

Scope2Screen realizes HTM as interactive multi-scale, tissue-aware lensing for $$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$7, 60-channel tissue images (Jessup et al., 2021). It exposes two nested lens modes: a Single-Cell Lens with small radius, on the order of $$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$8–$$r = \lfloor \gamma \cdot d_{\min}(M_t^{\mathrm{in}}) \rfloor,$$9, that pulls in the highest-resolution pyramid level, and a Tissue-Neighborhood Lens with larger radius, from several hundred $\gamma = 1/15$0 to more than $\gamma = 1/15$1, that blends in a lower-resolution layer to show tissue architecture without removing context. Dynamic scale switching computes the pyramid level according to

$\gamma = 1/15$2

and pre-fetches higher-detail tiles at a switch-over radius $\gamma = 1/15$3.

The rendering transition is mediated by three magnification styles: uniform magnifier, fisheye, and plateau. The plateau formulation uses an inner radius $\gamma = 1/15$4 at full magnification and an outer radius $\gamma = 1/15$5 where resolution is smoothly compressed into context. Context blending is given by

$\gamma = 1/15$6

which the paper states minimizes occlusion of contextual structures. Lens placement can also be segmentation-aware: when a released lens center lies within $\gamma = 1/15$7 pixels of a cell centroid, it snaps to that cell. Default magnifications are $\gamma = 1/15$8 for the single-cell lens and $\gamma = 1/15$9 for the neighborhood lens (Jessup et al., 2021).

The implementation combines OME-TIFF pyramids accessed via Zarr, WebGL2 fragment-shader compositing of up to 60 channels at approximately $I_r^k$00 per pixel, a server-side ball-tree over approximately $I_r^k$01 cell centroids supporting range queries in approximately $I_r^k$02–$I_r^k$03, and an integral-histogram-based HistoSearch whose viewport search takes approximately $I_r^k$04–$I_r^k$05 and whole-slide search approximately $I_r^k$06 to $I_r^k$07 (Jessup et al., 2021). Pure magnifier redraw remains below $I_r^k$08 for a Full-HD lens, lensing plus histogram update is approximately $I_r^k$09, and pan/zoom reaches $I_r^k$10–$I_r^k$11 within cached tiles.

The validation data also clarifies several limitations and misconceptions. In Scope2Screen, pathologists rated the normal and plateau magnifiers as “very intuitive” and “essential for preserving cell morphology,” whereas fisheye scored poorly, indicating that not every focus+context distortion is tissue-appropriate (Jessup et al., 2021). In HistoSelect, the current method does not generate textual rationales for why each group or patch was selected, validation is limited to TCGA and a private ovarian set, and future work is proposed on custom prompts, dynamic budgets, and jointly learned prompt embeddings (Huang et al., 28 Feb 2026). In EndoControlMag, current runtime is not adequate for standard live feedback rates, even though the method is described as robust and artifact-free under challenging surgical conditions (Wanga et al., 21 Jul 2025).

These limitations indicate that HTM is not merely a magnification heuristic. Its effectiveness depends on the fidelity of the structure-tracking or tissue-grouping stage, the continuity of the transition from focus to context, and the computational feasibility of deploying the hierarchy at the required temporal or spatial scale.