Coarse Guidance Network (CGN)

• Coarse Guidance Network (CGN) is a module that injects coarse spatial context into high-resolution patch features to enhance slide-level predictions in MIL frameworks.
• It remaps instance features to a coarse grid using field-of-view driven binning and processes them through a lightweight convolutional head to compute a guidance map.
• Empirical evaluations show that incorporating CGNs improves biomarker classification AUCs while maintaining low parameter and computational overhead.

A Coarse Guidance Network (CGN) is a module designed to learn and inject spatial contextual information at a coarser scale into high-magnification instance features within Multiple-Instance Learning (MIL) frameworks for whole-slide image (WSI) analysis. The CGN operates via grid-based remapping of instance features and a lightweight convolutional head to produce a coarse guidance map, which is then used to modulate the instance features before final attention-based aggregation. This approach enables progressive multi-scale context modeling in computational pathology tasks, offering a parameter-efficient mechanism for slide-level prediction enhancement while maintaining computational tractability (Wu et al., 2 Feb 2026).

1. Architectural Overview

The CGN processes high-magnification patch features $H \in \mathbb{R}^{N \times D}$ and their normalized spatial coordinates $(x'_n, y'_n)\in[0,1]^2$. Its core workflow includes three sequential steps:

1. Grid-based Remapping: High-magnification features are aggregated into a 3D coarse feature map $M \in \mathbb{R}^{D \times H' \times W'}$ based on spatial bin assignments determined by a selectable field-of-view (FOV) parameter.
2. Convolutional Guidance Head: $M$ is processed by two 3×3 convolutions with ReLU activations and a 1×1 convolution with Sigmoid activation to yield the coarse guidance map $P \in \mathbb{R}^{1 \times H' \times W'}$.
3. Patch-level Gating: $P$ is flattened and indexed to obtain $M_A \in \mathbb{R}^{N}$, which gates each corresponding row in $H$, resulting in modulated features $H_k = H \odot M_A$.

The diagrammatic ASCII representation is:

$P$5

2. Grid-based Remapping

Instance features and coordinates are mapped to a coarse grid via field-of-view driven binning. For each instance $n$ with normalized coordinates $(x'_n, y'_n)\in[0,1]^2$0, the grid cell assignment $(x'_n, y'_n)\in[0,1]^2$1 is determined as: $(x'_n, y'_n)\in[0,1]^2$2 where $(x'_n, y'_n)\in[0,1]^2$3, $(x'_n, y'_n)\in[0,1]^2$4, with $(x'_n, y'_n)\in[0,1]^2$5 the selected FOV.

The feature map $(x'_n, y'_n)\in[0,1]^2$6 is computed by averaging all high-magnification vectors falling into each bin: $(x'_n, y'_n)\in[0,1]^2$7 where $(x'_n, y'_n)\in[0,1]^2$8 collects all instances assigned to grid cell $(x'_n, y'_n)\in[0,1]^2$9. In vectorized notation,

$M \in \mathbb{R}^{D \times H' \times W'}$0

followed by reshaping $M \in \mathbb{R}^{D \times H' \times W'}$1 to $M \in \mathbb{R}^{D \times H' \times W'}$2.

3. Convolutional Guidance Computation

After remapping, $M \in \mathbb{R}^{D \times H' \times W'}$3 is passed through three sequential convolutions: $M \in \mathbb{R}^{D \times H' \times W'}$4

$M \in \mathbb{R}^{D \times H' \times W'}$5

$M \in \mathbb{R}^{D \times H' \times W'}$6

Here, $M \in \mathbb{R}^{D \times H' \times W'}$7 is used as the hidden channel width for all CGN blocks. No self-attention or Transformer module is included; the head is purely convolutional.

$M \in \mathbb{R}^{D \times H' \times W'}$8 is flattened to length $M \in \mathbb{R}^{D \times H' \times W'}$9, and each instance $M$0 gathers its coarse guidance value $M$1 according to its assigned index. The final gated features are $M$2.

4. Integration with Attention-based MIL

In standard attention-based MIL settings, instance embeddings $M$3 propagate through an attention aggregator $M$4 to yield slide-level predictions: $M$5 Installing a CGN at scale $M$6 updates $M$7 as: $M$8

Stacking multiple CGNs (for example, at FOVs $M$9) results in a progressive series of residual updates: $P \in \mathbb{R}^{1 \times H' \times W'}$0 The final $P \in \mathbb{R}^{1 \times H' \times W'}$1 is then input to attention modules such as ABMIL, DSMIL, CLAM-SB, or CLAM-MB, which conduct the slide-level aggregation.

5. Training Protocol and Hyperparameters

Training details for CGN-based models are as follows:

• Losses: Biomarker tasks (ER, PR, HER2 status) use cross-entropy loss. Prognosis tasks (CRC Surv) use a negative log-likelihood loss (NLLSurvLoss) that combines censored and uncensored terms:

$P \in \mathbb{R}^{1 \times H' \times W'}$2

$P \in \mathbb{R}^{1 \times H' \times W'}$3

$P \in \mathbb{R}^{1 \times H' \times W'}$4

• Optimizer: AdamW, learning rate $P \in \mathbb{R}^{1 \times H' \times W'}$5, cosine-decay scheduler.
• Early stopping: patience = 10.
• Epochs: maximum 150.
• FOV choices: At 20×, e.g., $P \in \mathbb{R}^{1 \times H' \times W'}$6 pixels (providing 3 CGNs).
• Hidden channels: $P \in \mathbb{R}^{1 \times H' \times W'}$7 per CGN.
• Parameter and compute cost: Each CGN adds approximately $P \in \mathbb{R}^{1 \times H' \times W'}$8M parameters per scale.

6. Empirical Performance and Ablation Results

Empirical studies isolating the CGN demonstrate a consistent benefit on multiple biomarker classification tasks. For instance, a single CGN (FOV=1536) added to ABMIL (using CONCH features) produces:

System	ER AUC (%)	PR AUC (%)	HER2 AUC (%)	Params (M)	FLOPs (G)
ABMIL w/o CGN (single-scale 20×)	87.22	84.14	80.06	—	—
ABMIL + single CGN (FOV=1536)	88.92	84.76	80.84	~1.51	~17.7
ABMIL + three CGNs ([1536,2048,3072])	89.76	85.24	82.86	~2.18	~17.7
ABMIL + five CGNs ([1024,1536,2048,2560,3072])	91.42	84.18	84.62	—	—

Adding at least one CGN leads to a clear increase in slide-level AUC—e.g., gains of +1.70pp (ER), +0.62pp (PR), and +0.78pp (HER2) for a single scale. Stacking multiple CGNs for progressive multi-scale guidance further improves performance (e.g., +4.20pp for ER, +4.56pp for HER2). CGNs achieve these gains at reduced parameter and compute cost relative to methods such as concatenation or cross-scale attention schemes ($P \in \mathbb{R}^{1 \times H' \times W'}$9M parameters/$P$0G FLOPs for CGN vs. $P$1M/$P$2G for cross-scale alternatives), while delivering larger accuracy improvements (+3.6pp ER, +4.05pp HER2).

7. Summary of Properties

A CGN remaps high-magnification features to a spatially coarse grid, applies a three-layer convolutional head to compute a coarse attention map, reprojects this map back to the patch level to gate the D-dimensional features, and is trained end-to-end via the same MIL objectives. Each CGN block is lightweight (requiring $P$3 hidden channels, $P$4M parameters per scale), incurs minimal additional computation, and has been shown to consistently improve slide-level prediction performance in clinical biomarker and prognosis tasks (Wu et al., 2 Feb 2026).