pith. sign in

arxiv: 2503.24135 · v4 · submitted 2025-03-31 · 💻 cs.CV

PixelCAM: Pixel Class Activation Mapping for Histology Image Classification and ROI Localization

Alexis Guichemerre , Soufiane Belharbi , Mohammadhadi Shateri , Luke McCaffrey , Eric Granger This is my paper

Pith reviewed 2026-05-22 22:13 UTC · model grok-4.3

classification 💻 cs.CV
keywords weakly supervised object localizationhistology image classificationpixel class activation mappingROI localizationmulti-task learningclass activation mappingpseudo-labelspixel-wise classification
0
0 comments X

The pith

PixelCAM enables joint training of classification and localization by adding a pixel-wise classifier in the shared encoder feature space.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes PixelCAM to overcome limitations in weakly supervised object localization for histology images. Standard single-step CAM methods suffer from under- or over-activation and asynchronous convergence between tasks, while two-step methods are limited by a frozen classifier. PixelCAM trains a foreground/background pixel-wise classifier alongside the image classifier using the shared encoder's pixel features and partial cross-entropy loss on pseudo-labels from a pretrained WSOL model. Both tasks are optimized simultaneously with standard gradient descent, and the pixel classifier integrates into CNN or transformer architectures without changes. A sympathetic reader would care because the approach promises more accurate ROI localization using only cheap image-level labels rather than costly pixel annotations.

Core claim

PixelCAM is a foreground/background pixel-wise classifier in the pixel-feature space of an image encoder shared with classification. It is trained with partial-cross entropy on pixel pseudo-labels collected from a pretrained WSOL model. Both the image and pixel-wise classifiers are trained simultaneously using standard gradient descent. This multi-task setup addresses the asynchronous convergence problem and improves ROI localization in histology images while handling out-of-distribution data better than prior single- or two-step WSOL strategies. The pixel classifier requires no architecture modifications for integration.

What carries the argument

PixelCAM, a foreground/background pixel-wise classifier in the pixel-feature space of a shared image encoder, trained with partial-cross entropy on pseudo-labels from a pretrained WSOL model.

If this is right

  • Simultaneous training resolves asynchronous convergence between the classification and localization tasks.
  • Training in pixel-feature space supports accurate foreground/background delineation and improved ROI localization.
  • The approach yields better results on out-of-distribution histology datasets than standard WSOL methods.
  • PixelCAM integrates directly into CNN- and transformer-based models without any architecture changes.
  • Partial-cross entropy enables effective use of the collected pixel pseudo-labels during joint optimization.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The method could support iterative self-training by using the refined pixel classifier to generate improved pseudo-labels for subsequent rounds.
  • The shared-encoder design suggests the learned features could transfer to related tasks such as full segmentation with little extra supervision.
  • Applying the same pixel-classifier idea to other imaging domains with scarce localization cues might reduce annotation costs beyond histology.
  • Robustness checks on pseudo-label quality from different base WSOL models would clarify how sensitive final performance is to the initial label source.

Load-bearing premise

The pseudo-labels collected from a pretrained WSOL model are sufficiently accurate to train the pixel classifier to learn discriminant features and accurate foreground/background delineation.

What would settle it

If adding the PixelCAM pixel classifier produces no gain or a loss in localization metrics such as pixel accuracy or IoU on a held-out histology test set compared to the base pretrained WSOL model alone, the benefit of the joint training procedure would be falsified.

read the original abstract

Weakly supervised object localization (WSOL) methods allow training models to classify images and localize ROIs. WSOL only requires low-cost image-class annotations yet provides a visually interpretable classifier. Standard WSOL methods rely on class activation mapping (CAM) methods to produce spatial localization maps according to a single- or two-step strategy. While both strategies have made significant progress, they still face several limitations with histology images. Single-step methods can easily result in under- or over-activation due to the limited visual ROI saliency in histology images and scarce localization cues. They also face the well-known issue of asynchronous convergence between classification and localization tasks. The two-step approach is sub-optimal because it is constrained to a frozen classifier, limiting the capacity for localization. Moreover, these methods also struggle when applied to out-of-distribution (OOD) datasets. In this paper, a multi-task approach for WSOL is introduced for simultaneous training of both tasks to address the asynchronous convergence problem. In particular, localization is performed in the pixel-feature space of an image encoder that is shared with classification. This allows learning discriminant features and accurate delineation of foreground/background regions to support ROI localization and image classification. We propose PixelCAM, a cost-effective foreground/background pixel-wise classifier in the pixel-feature space that allows for spatial object localization. Using partial-cross entropy, PixelCAM is trained using pixel pseudo-labels collected from a pretrained WSOL model. Both image and pixel-wise classifiers are trained simultaneously using standard gradient descent. In addition, our pixel classifier can easily be integrated into CNN- and transformer-based architectures without any modifications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper proposes PixelCAM, a foreground/background pixel-wise classifier operating in the pixel-feature space of a shared image encoder. PixelCAM is trained via partial cross-entropy on pixel pseudo-labels generated once by a pretrained WSOL model, while the image classifier and PixelCAM are updated simultaneously by gradient descent. The method is presented as a multi-task WSOL approach that addresses asynchronous convergence between classification and localization, improves ROI localization on histology images with low visual saliency, and handles OOD datasets better than standard single-step or two-step CAM-based WSOL strategies. It claims easy integration into CNN- and transformer-based architectures without modifications.

Significance. If the central claims were empirically validated, the joint optimization framework could offer a practical alternative to frozen two-step WSOL pipelines by allowing the localization signal to influence the shared encoder during training. The use of a lightweight pixel classifier in feature space and the partial-cross-entropy formulation are conceptually straightforward extensions of existing WSOL ideas. However, the absence of any reported experiments, ablations, or quantitative comparisons means the practical significance for histology imaging or OOD generalization cannot be assessed from the manuscript.

major comments (2)
  1. [Abstract] Abstract: the claims that simultaneous training 'addresses the asynchronous convergence problem' and yields 'better ROI localization' and improved OOD performance are presented without any experimental results, ablation studies, or quantitative metrics. This absence makes it impossible to evaluate whether the proposed joint optimization actually delivers the stated benefits over single-step or two-step baselines.
  2. [Abstract] PixelCAM training procedure (described in abstract): the method depends on the assumption that pseudo-labels collected from a pretrained WSOL model are sufficiently accurate for the pixel classifier to learn discriminant foreground/background features. The abstract itself states that standard WSOL methods suffer from under-/over-activation on histology images; no analysis, robustness argument, or derivation is supplied showing that the pixel-wise classifier can correct or improve upon errors in these fixed pseudo-labels rather than propagating them.
minor comments (1)
  1. Notation for the partial cross-entropy loss and the precise mechanism for generating and thresholding the pixel pseudo-labels should be defined explicitly, ideally with a short algorithmic description or equation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the careful review and constructive comments. We agree that the abstract overstates benefits without supporting evidence and that the pseudo-label assumption requires further justification. We will revise the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the claims that simultaneous training 'addresses the asynchronous convergence problem' and yields 'better ROI localization' and improved OOD performance are presented without any experimental results, ablation studies, or quantitative metrics. This absence makes it impossible to evaluate whether the proposed joint optimization actually delivers the stated benefits over single-step or two-step baselines.

    Authors: We agree that the abstract presents claims without accompanying experimental support. The submitted manuscript describes the proposed method but does not include the quantitative results, ablations, or comparisons needed to substantiate those claims. We will revise the abstract to describe the intended contributions without asserting unverified performance gains and will add the required experimental section with results on histology datasets and OOD evaluation in the revised version. revision: yes

  2. Referee: [Abstract] PixelCAM training procedure (described in abstract): the method depends on the assumption that pseudo-labels collected from a pretrained WSOL model are sufficiently accurate for the pixel classifier to learn discriminant foreground/background features. The abstract itself states that standard WSOL methods suffer from under-/over-activation on histology images; no analysis, robustness argument, or derivation is supplied showing that the pixel-wise classifier can correct or improve upon errors in these fixed pseudo-labels rather than propagating them.

    Authors: We agree that the manuscript provides no analysis or argument addressing whether the pixel classifier can improve upon errors in the fixed pseudo-labels. We will add a dedicated paragraph in the method or discussion section that examines this assumption, including a qualitative argument on how joint optimization of the shared encoder may allow refinement of features and a plan for an ablation study on pseudo-label noise in the revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper's central construction uses fixed pseudo-labels generated once by an external pretrained WSOL model to supervise a new PixelCAM pixel classifier in shared feature space, with simultaneous gradient updates on both tasks. This is a standard multi-task setup with an independent component; no equations or steps reduce by construction to the inputs, no self-citations are load-bearing for the core claim, and no predictions are statistically forced from fitted parameters. The derivation remains self-contained against the stated external WSOL source.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 1 invented entities

Based on abstract only; the central claim rests on domain assumptions about pseudo-label quality and shared feature space utility. No explicit free parameters or invented physical entities are described.

axioms (2)
  • domain assumption The feature space of the image encoder contains sufficient information for both image-level classification and pixel-level foreground/background separation.
    Invoked when localization is performed in the pixel-feature space of the shared encoder.
  • domain assumption Pixel pseudo-labels from a pretrained WSOL model provide reliable supervision for training the pixel classifier.
    Stated as the source of supervision for PixelCAM training with partial-cross entropy.
invented entities (1)
  • PixelCAM pixel-wise classifier no independent evidence
    purpose: Foreground/background classification in pixel-feature space to support ROI localization
    New component introduced for the multi-task WSOL approach.

pith-pipeline@v0.9.0 · 5837 in / 1332 out tokens · 56023 ms · 2026-05-22T22:13:12.276526+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.