About this case study

Clinicians require transparent AI systems before trusting automated diagnostic predictions. This project implements a Grad-CAM++ based explainability pipeline on a ResNet-50 backbone, trained on 112,000 CheXpert chest X-ray images across 14 pathology classes. The system generates visual attention maps that align with radiologist annotations, enabling clinicians to verify that the model focuses on diagnostically relevant regions before accepting its classification output.

Deep Learning Grad-CAM++ ResNet-50 Medical Imaging Explainable AI Python PyTorch

Key Metrics

112k Images Trained
89% Alignment
4.3/5 Trust Score
14 Pathologies

Interactive Explainability Prototype

Toggle visualization modes to explore how the model interprets chest X-rays.

🩹

Chest X-Ray Explainability Viewer

Grad-CAM++ Attention Visualization
Bone Structure
High Attention
Medium Attention
Low Attention

The Problem

Deep learning models for medical image classification have reached near-radiologist-level accuracy, yet their adoption in clinical workflows remains limited. The core issue is trust: when a convolutional neural network flags a chest X-ray as showing cardiomegaly, the clinician has no way to verify whether the model examined the cardiac silhouette or simply memorized spurious correlations from imaging artifacts. Black-box predictions create liability issues in healthcare settings, where every diagnostic decision must be explainable and auditable. Regulatory bodies increasingly require interpretability as a prerequisite for clinical deployment of AI systems.

The Approach

We implemented Grad-CAM++ on a ResNet-50 backbone pre-trained on ImageNet and fine-tuned on 112,000 CheXpert chest X-ray images spanning 14 pathology classes. Grad-CAM++ extends the original Grad-CAM by using a weighted combination of positive partial derivatives of the final convolutional layer, producing finer-grained attention maps that better localize multiple instances of the same pathology within a single image. We built a saliency alignment metric that quantitatively compares model attention heatmaps against bounding-box annotations provided by board-certified radiologists. The pipeline generates three complementary visualizations: Grad-CAM++ heatmap overlays, gradient-based saliency maps, and thresholded attention region outlines, giving clinicians multiple perspectives on the model's reasoning process.

The Results

The Grad-CAM++ pipeline achieved 89% saliency alignment with radiologist annotations across the CheXpert validation set, meaning the model's high-attention regions overlapped substantially with the anatomical areas radiologists identified as diagnostically relevant. In a user study with 23 clinicians, the trust score for AI-assisted diagnosis improved from 2.1 to 4.3 on a 5-point Likert scale after providing the explainability visualizations. The system correctly highlighted the cardiac silhouette for cardiomegaly cases, the costophrenic angles for pleural effusion, and the perihilar regions for pneumonia. Clinicians reported that the attention maps reduced their diagnostic review time by approximately 35% while maintaining their confidence in the final classification.