[Paper Review] Pulmonary Disease Classification Using Globally Correlated Maximum Likelihood: an Auxiliary Attention mechanism for Convolutional Neural Networks

Edward Verenich, Tobías Martín+3 more|arXiv (Cornell University)|Sep 1, 2021

COVID-19 diagnosis using AI4 citations

TL;DR

This paper proposes a novel auxiliary global attention mechanism, Globally Correlated Maximum Likelihood (GCML), that enhances convolutional neural networks (CNNs) for pulmonary disease classification by capturing global spatial correlations between salient features without sacrificing translation invariance or equivariance. GCML improves CNN sensitivity on pneumonia classification by 2.6 percentage points (93.7% vs. 91.1%) with 75% confidence, enabling better discrimination of similar diseases like viral pneumonia and COVID-19 using spatial context.

ABSTRACT

Convolutional neural networks (CNN) are now being widely used for classifying and detecting pulmonary abnormalities in chest radiographs. Two complementary generalization properties of CNNs, translation invariance and equivariance, are particularly useful in detecting manifested abnormalities associated with pulmonary disease, regardless of their spatial locations within the image. However, these properties also come with the loss of exact spatial information and global relative positions of abnormalities detected in local regions. Global relative positions of such abnormalities may help distinguish similar conditions, such as COVID-19 and viral pneumonia. In such instances, a global attention mechanism is needed, which CNNs do not support in their traditional architectures that aim for generalization afforded by translation invariance and equivariance. Vision Transformers provide a global attention mechanism, but lack translation invariance and equivariance, requiring significantly more training data samples to match generalization of CNNs. To address the loss of spatial information and global relations between features, while preserving the inductive biases of CNNs, we present a novel technique that serves as an auxiliary attention mechanism to existing CNN architectures, in order to extract global correlations between salient features.

Motivation & Objective

• To address the limitation of standard CNNs in capturing global spatial relationships between pulmonary abnormalities, which hinders discrimination of similar diseases like viral pneumonia and COVID-19.
• To preserve the inductive biases of CNNs—translation invariance and equivariance—while enabling global feature correlation awareness.
• To reduce data dependency compared to Vision Transformers by enabling effective generalization in data-starved medical imaging environments.
• To develop an auxiliary attention mechanism that operates on class activation maps rather than raw image tokens, improving spatial reasoning without altering the CNN backbone.

Proposed method

• GCML applies a global attention mechanism to class activation maps (CAMs) generated by the CNN, rather than to image patches or tokens, to preserve CNN inductive biases.
• It models the conditional probability distribution of salient features across the image using a thresholded binomial approximation (via τ parameter), learning global spatial correlations.
• The attention mechanism computes query, key, and value matrices from CAMs, enabling the network to weigh feature importance based on global interrelations.
• The method avoids image tokenization, maintaining the CNN’s spatial invariance while introducing discriminative global context for classification.
• GCML is implemented as a lightweight, differentiable module that can be added to existing CNN architectures without retraining the backbone.
• The model is trained end-to-end using a globally correlated maximum likelihood objective to optimize for spatially coherent feature representations.

Experimental results

Research questions

• RQ1 Can a global attention mechanism be integrated into standard CNNs to improve their ability to distinguish pulmonary diseases with similar features but different spatial distributions?
• RQ2 Does the proposed GCML mechanism preserve the translation invariance and equivariance of CNNs while enabling global spatial reasoning?
• RQ3 Can GCML achieve performance comparable to Vision Transformers with significantly fewer training samples, particularly in data-scarce medical imaging settings?
• RQ4 To what extent does incorporating global spatial context improve sensitivity in classifying viral pneumonia versus other similar conditions?

Key findings

• GCML achieved a test accuracy of 93.7% on pneumonia classification, compared to 91.1% for the standard CNN, representing a 2.6 percentage point improvement.
• The hypothesis test confirmed that GCML outperforms the standard CNN with 75% confidence (Z = -1.1607), indicating statistical significance in improved accuracy.
• The method improved sensitivity by leveraging global spatial correlations between salient features, enabling better discrimination of diseases with overlapping visual patterns.
• GCML maintains the inductive biases of CNNs—translation invariance and equivariance—while adding global attention, unlike Vision Transformers which lack these properties.
• The approach reduces data dependency compared to Vision Transformers, as GCML generalizes effectively even in data-starved environments.
• The authors released all code, data, and models for reproducibility and further research, supporting community-driven advancement in medical AI.