TLDR: This research introduces an explainable deep learning framework for detecting brain tumors in MRI scans and pneumonia in chest X-rays. Using ResNet50 and DenseNet121, the study found DenseNet121 consistently achieved higher accuracy. Crucially, by integrating Grad-CAM, the researchers showed that DenseNet121 provided more clinically relevant explanations by focusing precisely on pathological regions, unlike ResNet50 which sometimes highlighted irrelevant areas. This highlights the importance of explainable AI in building trust and facilitating the adoption of AI in clinical diagnostics.
Deep Learning (DL) has shown immense promise in revolutionizing medical imaging diagnostics, offering the potential to significantly improve how diseases like brain tumors and pneumonia are detected. However, a major hurdle to widespread clinical adoption has been the ‘black-box’ nature of most deep learning models – clinicians often find it difficult to trust AI predictions without understanding how they were made. This lack of interpretability can hinder confidence and integration into real-world healthcare settings.
A recent study addresses this critical challenge by presenting an explainable deep learning framework for two vital medical diagnostic tasks: detecting brain tumors in MRI scans and identifying pneumonia in chest X-ray images. The research, conducted by Sai Teja Erukude, Viswa Chaitanya Marella, and Suhasnadh Reddy Veluru, leverages two prominent Convolutional Neural Networks (CNNs), ResNet50 and DenseNet121, and integrates an explainability technique called Grad-CAM.
The Models and Data
The researchers trained ResNet50 and DenseNet121 on publicly available datasets from Kaggle. For brain tumor detection, they used a dataset of over 7,000 brain MRI images, categorized into ‘tumor’ and ‘no tumor’. For pneumonia detection, a dataset of nearly 6,000 pediatric chest X-ray images, classified as ‘Pneumonia’ or ‘Normal’, was utilized. Both models were initialized with weights pre-trained on ImageNet, a common practice in transfer learning, and then fine-tuned on these medical imaging datasets.
Performance and Key Findings
Both ResNet50 and DenseNet121 demonstrated strong classification performance across both tasks. However, DenseNet121 consistently outperformed ResNet50 in terms of accuracy: achieving 94.3% accuracy for brain tumor detection compared to ResNet50’s 92.5%, and 89.1% accuracy for pneumonia detection versus ResNet50’s 84.4%. DenseNet121 also showed superior performance in other metrics like AUC and F1-score, indicating a more robust and balanced predictive capability.
The Power of Explainable AI with Grad-CAM
The most significant contribution of this research lies in its integration of Gradient-weighted Class Activation Mapping (Grad-CAM) to provide visual explanations for the models’ decisions. Grad-CAM generates heatmaps that are overlaid onto the original medical images, highlighting the specific regions that were most influential in the model’s prediction. This allows human experts to visually understand where the AI is focusing its attention.
The Grad-CAM visualizations revealed crucial differences between the two architectures. While both models produced accurate results, DenseNet121 consistently focused its attention precisely on the core pathological regions – for instance, directly on the tumor area in MRI scans or on the opacities in lung fields indicative of pneumonia. In contrast, ResNet50, though accurate, sometimes scattered its attention to peripheral or non-pathological areas, such as the skull in MRI images or the heart and ribs in chest X-rays. This suggests that DenseNet121’s dense connectivity allows it to home in on the most discriminative features, avoiding distractions from irrelevant parts of the image.
Also Read:
- Predicting Chemical Toxicity with Visual AI and Explainable Insights
- Assessing GPT-4o’s Ability to Detect Pneumonia from X-Rays
Building Trust in Clinical Settings
These findings underscore a vital point: in healthcare, accuracy alone is not sufficient. The ability to explain an AI model’s decision-making process is paramount for building trust among clinicians and facilitating the adoption of AI systems. A model that can visually demonstrate its focus on medically relevant areas will be more readily validated and trusted by human experts. The combination of high-performing models like DenseNet121 with intuitive explainability tools like Grad-CAM offers a promising pathway toward creating reliable, interpretable, and clinically useful diagnostic AI tools.
The researchers believe that this approach will form a foundation for increasing trust, improving clinical acceptance, and ultimately leading to the real-world deployment of AI systems in healthcare. Future work will explore expanding to new diseases, testing advanced XAI techniques, and conducting clinician-centered user studies to further realize the potential of explainable deep learning in improving patient care. You can read the full research paper here.
