AppendiGrade: An XAI-Enhanced Deep Learning Framework for Grading Appendicitis in Ultrasound with Gaussian Blur and Grad-CAM

AI can significantly improve diagnostic accuracy for serious conditions: The AppendiGrade system achieved 95.58% accuracy in detecting complicated appendicitis cases (perforation, abscess) from ultrasound images, compared to 69.21% with unoptimized models, demonstrating that deep learning can substantially enhance clinical decision-making for time-sensitive emergencies.

Explainability tools are essential for clinical AI adoption: The researchers used Grad-CAM (gradient-weighted class activation mapping) to create visual heatmaps showing which regions of ultrasound images the AI focused on when making diagnoses, making the "black box" decision-making process transparent and verifiable by radiologists before clinical use.

Data quality and preprocessing matter as much as the algorithm itself: Simply applying raw images to pre-trained models yielded mediocre results (69.21% accuracy), but systematic improvements including image preprocessing, hyperparameter tuning, and image sharpening boosted performance to 95.58%—highlighting that responsible AI requires careful attention to data preparation, not just sophisticated models.

AI can reduce radiation exposure and healthcare costs: By improving ultrasound-based diagnosis of appendicitis, this system supports safer, radiation-free diagnostic pathways compared to CT scans, potentially reducing both patient harm and healthcare expenses while maintaining diagnostic reliability.

Diverse model comparison strengthens validation: The research tested four different deep learning architectures (DenseNet201, InceptionV3, ConvNextTiny, VGG19) rather than relying on a single model, providing more robust evidence that the approach works reliably across different AI systems.