Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Gradient Attention Map Based Verification of Deep Convolutional Neural Networks with Application to X-ray Image Datasets

Omid Halimi Milani, Amanda Nikho, Lauren Mills, Marouane Tliba, Ahmet Enis Cetin, Mohammed H. Elnagar

TL;DR

The paper tackles the risk of misapplying deep CNNs in medical imaging by introducing a comprehensive verification framework. It combines Gradient Attention Map (GAM) similarity analysis, early-layer feature-map verification, and a garbage-class mechanism to explicitly reject out-of-distribution inputs. Key contributions include a seven-metric GAM comparison feeding a Random Forest, a parallel feature-map verification scheme, and a garbage-class strategy that enhances robustness and reduces misclassifications. Empirical results on SOS fusion staging and Class III malocclusion X-ray datasets show strong attention alignment, high discrimination between suitable and unsuitable models, and improved resilience to out-of-distribution data, supporting safer clinical deployment.

Abstract

Deep learning models have great potential in medical imaging, including orthodontics and skeletal maturity assessment. However, applying a model to data different from its training set can lead to unreliable predictions that may impact patient care. To address this, we propose a comprehensive verification framework that evaluates model suitability through multiple complementary strategies. First, we introduce a Gradient Attention Map (GAM)-based approach that analyzes attention patterns using Grad-CAM and compares them via similarity metrics such as IoU, Dice Similarity, SSIM, Cosine Similarity, Pearson Correlation, KL Divergence, and Wasserstein Distance. Second, we extend verification to early convolutional feature maps, capturing structural mis-alignments missed by attention alone. Finally, we incorporate an additional garbage class into the classification model to explicitly reject out-of-distribution inputs. Experimental results demonstrate that these combined methods effectively identify unsuitable models and inputs, promoting safer and more reliable deployment of deep learning in medical imaging.

Gradient Attention Map Based Verification of Deep Convolutional Neural Networks with Application to X-ray Image Datasets

TL;DR

The paper tackles the risk of misapplying deep CNNs in medical imaging by introducing a comprehensive verification framework. It combines Gradient Attention Map (GAM) similarity analysis, early-layer feature-map verification, and a garbage-class mechanism to explicitly reject out-of-distribution inputs. Key contributions include a seven-metric GAM comparison feeding a Random Forest, a parallel feature-map verification scheme, and a garbage-class strategy that enhances robustness and reduces misclassifications. Empirical results on SOS fusion staging and Class III malocclusion X-ray datasets show strong attention alignment, high discrimination between suitable and unsuitable models, and improved resilience to out-of-distribution data, supporting safer clinical deployment.

Abstract

Deep learning models have great potential in medical imaging, including orthodontics and skeletal maturity assessment. However, applying a model to data different from its training set can lead to unreliable predictions that may impact patient care. To address this, we propose a comprehensive verification framework that evaluates model suitability through multiple complementary strategies. First, we introduce a Gradient Attention Map (GAM)-based approach that analyzes attention patterns using Grad-CAM and compares them via similarity metrics such as IoU, Dice Similarity, SSIM, Cosine Similarity, Pearson Correlation, KL Divergence, and Wasserstein Distance. Second, we extend verification to early convolutional feature maps, capturing structural mis-alignments missed by attention alone. Finally, we incorporate an additional garbage class into the classification model to explicitly reject out-of-distribution inputs. Experimental results demonstrate that these combined methods effectively identify unsuitable models and inputs, promoting safer and more reliable deployment of deep learning in medical imaging.
Paper Structure (21 sections, 10 equations, 2 figures, 6 tables)

This paper contains 21 sections, 10 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. X-ray Image Datasets
  3. SOS Fusion Staging Dataset
  4. Class III Malocclusion X-ray Dataset
  5. Methodology
  6. Model Verification Using Grad-CAM Similarity Metrics
  7. Grad-CAM Computation
  8. Threshold-Based Binarization
  9. Spatial Overlap Metrics
  10. Perceptual Similarity Assessment
  11. Statistical Correlation Analysis
  12. Distributional Divergence Measures
  13. Integrating Similarity Metrics into the Verification Framework
  14. Random Forest-based Model Verification and Rejection Strategy
  15. Feature Map-Based Verification via Internal Representations (Method 2)
  16. ...and 6 more sections

Figures (2)

  • Figure 1: Grad-CAM visualization for a model trained and tested on the SOS dataset. The attention is well-aligned with the anatomical structures, highlighting relevant areas.
  • Figure 2: Visualization of the average Gradient Attention Maps (GAMs) for each model during training. The first row represents attention maps for models trained on the SOS dataset and evaluated on the SOS training set. The second row shows attention maps for models trained on the Class III dataset and tested on the SOS training set. The heatmaps illustrate where each model focuses when making predictions.