Feedback Attention Ladder CNN

• The paper presents FAL-CNN which enhances detection by generating spatial attention maps and class probabilities for guiding false positive correction.
• FAL-CNN employs a ladder-style architecture based on VGG-19, where symmetric feedforward and feedback branches use 1×1 convolutions to compute attention at multiple scales.
• Although preliminary F1 scores are lower than baseline FCOS, FAL-CNN is posited to improve detection robustness and interpretability through top–down modulation.

Feedback Attention Ladder CNN (FAL-CNN) is a classifier-and-attention module introduced by Broad et al. within the composite pipeline of “MIDOG 2025: Mitotic Figure Detection with Attention-Guided False Positive Correction” (Broad et al., 29 Aug 2025). In that system, FAL-CNN extends a Fully Convolutional One-Stage Object Detector (FCOS) for mitotic figure detection by producing spatial attention maps and a class probability that are consumed by a downstream fusion network trained to adjust FCOS bounding boxes. The stated objective is to reduce the false positive rate of the FCOS object detector, improve object-detection accuracy, and enhance generalisability; on the preliminary MIDOG 2025 evaluation dataset, the composite model achieved an F1 score of 0.655 (Broad et al., 29 Aug 2025).

1. Placement within the MIDOG 2025 pipeline

FAL-CNN is not presented as a standalone detector. It is one of four modules in a composite mitosis-detection pipeline: an FCOS one-stage detector, a Feedback Attention Ladder CNN classifier, a fusion network for False Positive Correction, and a mini-patch sampler (Broad et al., 29 Aug 2025).

Module	Role
FCOS one-stage detector	Produces initial bounding boxes
FAL-CNN classifier	Produces attention map and class probability
Fusion network	Generates box adjustments and score modification
Mini-patch sampler	Extracts 56×56 px mini-patches

The processing flow is summarized as: input image $\rightarrow$ FCOS $\rightarrow$ initial bboxes $\rightarrow$ sample 56×56 px mini-patches $\rightarrow$ FAL-CNN $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$ $\rightarrow$ fusion network $\rightarrow$ adjusted bboxes $\rightarrow$ final output (Broad et al., 29 Aug 2025). Within this arrangement, FAL-CNN operates on patch-centered cutouts rather than on the full image.

The abstract describes FAL-CNN as supporting “classification of normal versus abnormal mitotic figures,” whereas the architectural walkthrough specifies a two-way output producing a mitosis-versus-non-mitosis probability $p_{\rm mitosis}$ (Broad et al., 29 Aug 2025). The available description therefore uses both phrasings. What is unambiguous is that the FAL-CNN output is subsequently fused with FCOS confidence information for false-positive correction.

2. Feedforward backbone and ladder construction

The feedforward path of FAL-CNN is based on VGG-19, with five convolutional stages whose exact filter counts, sizes, and strides follow the VGG-19 archetype but are not fully enumerated in the paper (Broad et al., 29 Aug 2025). The architecture is described as follows: Conv Block 1 uses 3×3 convolutions with 64 channels and 2×2 max-pooling; Conv Block 2 uses 3×3 convolutions with 128 channels and 2×2 max-pooling; Conv Block 3 uses 3×3 convolutions with 256 channels and 2×2 max-pooling; Conv Blocks 4 and 5 use 3×3 convolutions with 512 channels, with max-pooling after Block 4. All convolutions use stride $=1$, padding to preserve spatial size, and ReLU activations.

The feedback branch is symmetrical and mirrors the five feedforward blocks in reverse. After each feedforward block $\rightarrow$0, features $\rightarrow$1 are passed through a corresponding feedback block to compute an attention map $\rightarrow$2. Each feedback block typically consists of a $\rightarrow$3 convolution with channels $\rightarrow$4 and bias $\rightarrow$5, followed by ReLU, optional up-sampling by a factor of $\rightarrow$6 via nearest-neighbour or deconvolution to match the spatial size of the previous block, and a sigmoid non-linearity to yield $\rightarrow$7 (Broad et al., 29 Aug 2025).

This ladder-style arrangement injects top-down context into each feedforward stage. The architectural summary explicitly identifies the hierarchical feedback attention ladder as one of the key innovations of the MIDOG 2025 system. A plausible implication is that the model is intended to combine local evidence from patch content with higher-level contextual modulation before final classification and box correction.

3. Attention computation and gated propagation

The core top-down attention mechanism is defined ladder level by ladder level. At level $\rightarrow$8, the top-down attention is computed as

$\rightarrow$9

$\rightarrow$0

The resulting gated feature is

$\rightarrow$1

Here $\rightarrow$2 denotes a $\rightarrow$3 convolution, $\rightarrow$4 denotes the element-wise sigmoid, and the attention map is broadcast along the channel axis before element-wise multiplication (Broad et al., 29 Aug 2025). The gated features $\rightarrow$5 then proceed to the next deeper convolutional block.

The ladder connection is described procedurally: at each encoder stage $\rightarrow$6, the feedforward feature $\rightarrow$7 is combined with the upsampled feedback from stage $\rightarrow$8. The provided pseudocode uses a reverse traversal from stage $\rightarrow$9 to stage $\rightarrow$0, computes $\rightarrow$1 with a sigmoid applied to a $\rightarrow$2 convolution, multiplies $\rightarrow$3 by $\rightarrow$4, and, when $\rightarrow$5, upsamples $\rightarrow$6 by scale $\rightarrow$7 before passing it into the next shallower feedforward block (Broad et al., 29 Aug 2025). The description is explicit that the mechanism is based on multiplication, referred to as “attention gating,” and not on concatenation or addition.

In functional terms, FAL-CNN therefore differs from generic skip-connected fusion schemes by using scalar spatial masks to modulate intermediate activations. This suggests a design in which higher-level structure selectively suppresses or preserves lower-level responses rather than merely aggregating them.

4. Classification head and false-positive correction

After the final gated feature map $\rightarrow$8, global pooling reduces spatial dimensions, and the classification head applies a two-way fully connected layer plus softmax to produce the mitosis-versus-non-mitosis probability $\rightarrow$9 (Broad et al., 29 Aug 2025). In the more detailed branch description, this is written as global average pooling to a vector $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$0, followed by a fully connected layer and softmax to obtain $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$1.

The downstream fusion network combines three inputs: spatial attention map(s), flattened to a small 2-D embedding; the FCOS box confidence score $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$2; and $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$3 from FAL-CNN (Broad et al., 29 Aug 2025). The network is a small multilayer perceptron composed of alternating linear layers and ReLU activations, with hidden dimensions not specified. It produces four regression heads for box-offset adjustments $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$4 and one score multiplier $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$5.

Given an FCOS-predicted bounding box $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$6 with score $$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$7, the adjusted box and score are described by standard anchor-free regression:

$$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$8

$$\rightarrow \{\text{attention map},\, p_{\rm mitosis}\}$$9

$\rightarrow$0

The source notes that the paper states only that the fusion network is trained to produce both coordinate adjustments and score multipliers; the exact formulation was not given, and the displayed equations follow common practice (Broad et al., 29 Aug 2025). Even with that caveat, the operational role of FAL-CNN is clear: it does not replace FCOS, but rather supplies attention-derived and classification-derived signals for post-detection correction.

5. Training specification, stated omissions, and empirical evaluation

The available description is explicit that the original text does not supply a complete loss-function formula or all hyperparameters (Broad et al., 29 Aug 2025). Learning rate, batch size, number of epochs, and data augmentation are not stated. The optimizer is also not stated, although the description notes that the paper says the relevant component was “trained in situ with FCOS and FAL-CNN frozen.”

Because of these omissions, the only detailed training objective provided is inferential rather than definitive. Based on the description, one can surmise a total loss of the form

$\rightarrow$1

Within that surmised formulation, $\rightarrow$2 is the baseline object-detection loss, $\rightarrow$3 is the two-class cross-entropy loss on $\rightarrow$4, $\rightarrow$5 is a Smooth-L1 or IoU loss on the predicted offsets, and $\rightarrow$6 is an optional attention-regularization term (Broad et al., 29 Aug 2025). These terms should therefore be read as a reconstruction of likely training practice, not as a verbatim specification from the paper.

The reported empirical results are more direct. On the preliminary MIDOG 2025 evaluation set, the composite model consisting of FCOS, FAL-CNN, and the fusion network achieved $\rightarrow$7, while the baseline FCOS trained on the same data achieved $\rightarrow$8 (Broad et al., 29 Aug 2025). Precision and recall were not individually reported. No ablation studies, such as FCOS plus FAL-CNN without the fusion network, and no per-component performance breakdowns were included.

A common misconception would be to treat FAL-CNN as having already demonstrated superior detection performance to baseline FCOS on the reported evaluation. The preliminary numbers do not support that reading. The authors instead note that, although the initial test performance was lower, they expect increased domain-shift robustness and improved explainability via the attention maps (Broad et al., 29 Aug 2025). As stated, those expected benefits remain prospective within the evidence provided.

6. Relation to earlier feedback-attention research

A broader antecedent for feedback-based internal attention is “Deep Networks with Internal Selective Attention through Feedback Connections,” which introduced the Deep Attention Selective Network (dasNet) (Stollenga et al., 2014). DasNet starts from a feedforward Maxout CNN, adds scalar gates $\rightarrow$9 at each convolutional layer and output map, and uses a learned policy network to generate top-down actions that modulate convolutional filter sensitivities over multiple processing passes. Its gated convolution is written as

$\rightarrow$0

with the next action computed from an observation vector built from average activations, pre-softmax activations, and current class probabilities. The policy is deterministic,

$\rightarrow$1

and feedback is trained through separable natural evolution strategies (SNES) in a million-dimensional parameter space (Stollenga et al., 2014).

DasNet reported improved classification performance over its baseline single Maxout CNN, reducing CIFAR-10 error from $\rightarrow$2 to $\rightarrow$3 and CIFAR-100 error from $\rightarrow$4 to $\rightarrow$5, with stable behavior up to nine passes and peak performance at $\rightarrow$6 passes (Stollenga et al., 2014). Its visual analyses further suggested that top-down gating can correct early mistakes by shifting emphasis across filters and layers.

FAL-CNN, as described in MIDOG 2025, is architecturally different. It uses a symmetrical feedback path and ladder-style spatial attention maps rather than a policy network that iteratively emits layer-wise scalar gates over multiple passes. Nonetheless, the comparison is informative because both systems center top-down modulation of internal representations. The dasNet paper’s concluding outlook explicitly suggests that, in a Feedback Attention Ladder CNN, one could introduce gate vectors at multiple rungs of a ladder hierarchy, extend the observation to include lateral reconstructions or error signals, integrate the feedback policy into weighting of skip-connections, and train the policy end-to-end with gradient estimates or hybrid RL/gradient methods (Stollenga et al., 2014). Those statements are proposals rather than part of the MIDOG 2025 implementation, but they situate FAL-CNN within a longer line of research on feedback-driven selective attention.

In that sense, FAL-CNN can be understood as a ladder-based realization of internal attention for a detection-correction pipeline. The MIDOG 2025 description assigns its principal significance to the combination of multi-scale top-down attention, patch-level classification, and a downstream box-correction network, while also leaving several implementation specifics unspecified (Broad et al., 29 Aug 2025).