Accepted at IEEE AIC 2026: Uncertainty-Aware Attention for Clinical AI

AI models in medical imaging are reaching human-level accuracy, but in hospital settings, a major obstacle remains: black-box predictions. Most deep learning models produce binary decisions without explaining why or indicating how confident they are.

In my upcoming paper accepted at the IEEE AIC 2026 Conference, I developed a framework from the Department of Information Technology at Government College of Engineering, Chhatrapati Sambhajinagar designed to bridge this clinical trust gap.

---

The Problem: Why High Accuracy Isn't Enough for Healthcare

Standard Convolutional Neural Networks (CNNs) output a diagnosis even when image quality is poor or signs of disease are subtle. A doctor cannot distinguish whether an AI model is making a careful, evidence-based prediction or simply making a high-confidence guess.

Missing a diagnosis (a false negative) in conditions like pediatric pneumonia can delay life-saving treatment. To make AI clinically useful, systems must fulfill two requirements:

1. Explainability: Highlighting where the model is looking.
2. Uncertainty Quantification: Knowing when to say "I am unsure — flag this for a human radiologist."

---

High-Level Approach: Dual Attention + MC Dropout

My framework combines two complementary mechanisms to address both explainability and reliability:

1. Dual Attention (Channel + Spatial)

• Channel Attention: Learns which feature representations (e.g., texture vs edge patterns) carry the highest diagnostic signal.
• Spatial Attention: Focuses the model's receptive field on lung regions while dimming out irrelevant background anatomy.

2. Monte Carlo Dropout for Clinical Triage

Rather than turning off dropout during inference, I kept dropout active to execute multiple stochastic passes per image.

• High variance across passes indicates high model uncertainty.
• By setting an uncertainty threshold at the 95th percentile, the system automatically flags doubtful cases for human review.

---

Key Results & Impact

Evaluated on 5,856 pediatric chest X-rays:

• 97.18% Overall Accuracy (with 96.73% Sensitivity).
• Clinical Triage Concentration: Flagging just 5.8% of total cases captured 43% of all actual model errors — allowing radiologists to focus their review where human expertise is needed most.
• Visual Interpretability: Grad-CAM heatmaps confirmed that attention maps focus strictly on lung opacities rather than background artifacts.

---

Looking Ahead

This work demonstrates that combining attention mechanisms with uncertainty estimation significantly improves the safety and interpretability of AI in healthcare.

Paper accepted for presentation at **IEEE AIC 2026. Full publication details and code repository will be released following the conference proceedings.

---

Written by Rugved Chandekar (rugved.me) · IEEE AIC 2026 Paper Author