Explainable AI-Based Pulmonary Nodule Classification from Computed Tomography Images with an End-to-End Clinical Decision Support Interface

Abstract Early and accurate detection of pulmonary nodules in computed tomography (CT) imaging is critical for reducing lung cancer mortality. Although deep learning-based computer-aided detection (CAD) systems have demonstrated promis-ing diagnostic performance, their widespread clinical adoption has been impeded by two persistent limitations: the absence of visual explainability and the lack of deployable end-to-end interfaces that integrate prediction, explanation, and clinical workflow. This paper presents an explainable, end-to-end CAD system for pulmonary nodule classification that addresses both shortcomings. The proposed system employs a transfer-learned ResNet-50 architecture, fine-tuned on 2,372 balanced CT patches extracted from the publicly available LUNA16 dataset. Gradient-weighted Class Activation Mapping (Grad-CAM) is integrated into the inference pipeline to generate per-prediction heatmaps that visually localize the discriminative regions driv-ing each classification decision. The complete system is deployed as a production-grade web application featuring a FastAPI backend and a React frontend, complete with authentication and database persistence. On a held-out test set of 356 samples, the proposed system achieves an accuracy of 90.45%, precision of 90.91%, recall of 89.89%, F1 score of 90.40%, and an AUC-ROC of 0.9615. A confidence-based abstention mechanism returns an “Inconclusive” verdict for predictions below a 60% probability threshold, reducing the risk of unreliable automated decisions. Rather than claiming state-of-the-art raw accuracy, this work’s principal contribution is demonstrating that competitive classification performance, real-time visual explainability, and a fully deployed, browser-accessible clinical interface can be integrated into a single reproducible, open-source system, contributing a practical baseline for future hospital validation studies.