<p>Chest radiology images are used to diagnose many diseases, such as pneumonia, COVID-19, fractures, cardiomegaly, and several similar diseases. X-ray images are useful tools for the early diagnosis of fatal diseases. To provide facilities for X-ray diagnosis in remote areas and to speed up the diagnosis process, automation of X-ray image diagnosis is required. Automatic X-ray diagnosis is progressing in small steps using machine learning and deep learning. One of the main reasons for the resistance to automation in the medical field is trust issues, one of the common causes of which is the black-box nature of deep learning models. We use the XAI (Explainable Artificial intelligence) approach to establish trust in artificial intelligence methods. Some approaches to XAI include CAM, Grad-CAM, SHAP, Grad-CAM++, LIME, and Score-CAM. In this study, we propose a Model-specific Ensemble Local Explainable Artificial Intelligence (MEL-XAI). MEL-XAI aims to provide a confident result for the visual explanation by utilizing the power of "The Wisdom of Crowds" by ensembling Grad-CAM, Grad-CAM++, and Score-CAM. The approach processes the final layer features of the CNN, including the map, gradient, and weight, and then applies these features to three distinct explainable methods: Grad-CAM, Grad-CAM++, and Score-CAM. Each pixel in the heatmap represents a foundation color, with the coldest color(blue) denoting the lowest numerical value and the hottest color (red) denoting the highest. Hot colors are used to represent the most important features of the image; conversely, light colors represent the least important features of the image. The maximum pixel voting technique combines the results of Grad-CAM, Grad-CAM++, and Score-CAM to generate the MEL-XAI heatmap. Every CNN base architecture can utilize the model-specific MEL-XAI approach, as demonstrated in the explanation using VGG16 and ResNet50. The two datasets utilized were the general dataset ImageNet and a medical dataset for Pneumonia Chest X-ray(PCX) classification. In the MEL-XAI explanation with the VGG16 model pneumonia classification data, the % average drop was 38.84, and with the ImageNet dataset, the % average drop was 69.49. The MEL-XAI achieved an execution time of 924.35 ms (milliseconds) for a single image. The MEL-XAI, along with the chest X-ray dataset and ResNet50 model, outperformed on the scale of a % increase in confidence, with a value of 57.32. For Pneumonia chest X-ray, the mean vote from the radiologist panel was 73.33%. After the paper publication the code is available at <a href="https://github.com/nandanii/MEL-XAI.git">https://github.com/nandanii/MEL-XAI.git</a>.</p>

MEL-XAI Model-Specific Ensemble Local Explainable Artificial Intelligence Approach

期刊论文
发表日期：

2026年3月24日

查看全文

Nandani Sharma,
Sandeep Chaurasia

摘要

Chest radiology images are used to diagnose many diseases, such as pneumonia, COVID-19, fractures, cardiomegaly, and several similar diseases. X-ray images are useful tools for the early diagnosis of fatal diseases. To provide facilities for X-ray diagnosis in remote areas and to speed up the diagnosis process, automation of X-ray image diagnosis is required. Automatic X-ray diagnosis is progressing in small steps using machine learning and deep learning. One of the main reasons for the resistance to automation in the medical field is trust issues, one of the common causes of which is the black-box nature of deep learning models. We use the XAI (Explainable Artificial intelligence) approach to establish trust in artificial intelligence methods. Some approaches to XAI include CAM, Grad-CAM, SHAP, Grad-CAM++, LIME, and Score-CAM. In this study, we propose a Model-specific Ensemble Local Explainable Artificial Intelligence (MEL-XAI). MEL-XAI aims to provide a confident result for the visual explanation by utilizing the power of "The Wisdom of Crowds" by ensembling Grad-CAM, Grad-CAM++, and Score-CAM. The approach processes the final layer features of the CNN, including the map, gradient, and weight, and then applies these features to three distinct explainable methods: Grad-CAM, Grad-CAM++, and Score-CAM. Each pixel in the heatmap represents a foundation color, with the coldest color(blue) denoting the lowest numerical value and the hottest color (red) denoting the highest. Hot colors are used to represent the most important features of the image; conversely, light colors represent the least important features of the image. The maximum pixel voting technique combines the results of Grad-CAM, Grad-CAM++, and Score-CAM to generate the MEL-XAI heatmap. Every CNN base architecture can utilize the model-specific MEL-XAI approach, as demonstrated in the explanation using VGG16 and ResNet50. The two datasets utilized were the general dataset ImageNet and a medical dataset for Pneumonia Chest X-ray(PCX) classification. In the MEL-XAI explanation with the VGG16 model pneumonia classification data, the % average drop was 38.84, and with the ImageNet dataset, the % average drop was 69.49. The MEL-XAI achieved an execution time of 924.35 ms (milliseconds) for a single image. The MEL-XAI, along with the chest X-ray dataset and ResNet50 model, outperformed on the scale of a % increase in confidence, with a value of 57.32. For Pneumonia chest X-ray, the mean vote from the radiologist panel was 73.33%. After the paper publication the code is available at https://github.com/nandanii/MEL-XAI.git.