Explainable AI for gastrointestinal lesion surveillance and precision targeted drug delivery

TL;DR

This paper introduces an AI-powered ingestible device that can detect gastrointestinal diseases and deliver targeted drugs, with a focus on transparency and safety.

Contribution

The novel contribution is an AI-assisted IoBNT system combining GI imaging, explainable AI, and secure drug delivery with evaluation on diverse disease classes.

Findings

The system achieves 91.4% classification accuracy on the HyperKvasir dataset.

Explainable AI methods improve model transparency and validation.

Drug transport is modeled with pharmacokinetic frameworks and uncertainty analysis.

Abstract

The Internet of Bio-NanoThings (IoBNT) promises revolutionary healthcare applications, particularly in targeted drug delivery. However, major challenges remain including safe nanodevice design, monitoring their behavior in biological environments and enabling reliable communication with external control systems. This work proposes an AI-assisted IoBNT architecture that combines gastrointestinal (GI) imaging with intelligent therapeutic supervision. A wireless ingestible imaging device (WIID) captures GI images, while an Artificial Intelligence Ciphered Link (AICL) analyzes them using convolutional neural networks (CNNs) trained with supervised contrastive learning and cost-sensitive fine-tuning. Unlike prior studies focused solely on tumors, our system is evaluated on the HyperKvasir dataset covering 25 GI disease classes, including neoplastic and inflammatory conditions. Explainable AI methods (GradCAM family) are employed with quantitative validation to improve model transparency. Drug transport and release are modeled using a multi-compartment pharmacokinetic framework with uncertainty analysis. Security protections using Quadratic Map Privacy Algorithm (QMPA), threat modeling and failsafe dosing limits are incorporated to enhance clinical safety. The system achieves 91.4% classification accuracy (weighted F1 = 0.91) on HyperKvasir, with stronger performance in neoplastic classes and lower accuracy in rare categories, emphasizing the importance of class-balanced evaluation. These results demonstrate the feasibility of integrating AI-based disease detection with controlled drug delivery, representing a step toward closed-loop, adaptive IoBNT therapeutics.

Figures