Building Trust with Explainable AI in Lung Disease Detection Using Deep Learning

Abstract

The primary aim of this research is to design an XAI system for lung disease diagnosis that achieves the highest degree of accuracy possible in a more user-friendly manner than current systems. The EfficientB5 Network model is proposed for implementation as the main architecture. It has got the promise to achieve high accuracy rates up to 99.13%, but with more confidence, it could classify three types of lung diagnoses: Benign, Malignant, and Normal. For the augmentation of trust and usability in clinical practice, Grad-CAM visualization techniques based on Explainable AI will be utilized. The methodology highlights the class-specific probabilities and regional contributions as interpretable justification from the patient's or clinician's viewpoint for understanding the prediction outcome. While lung cancer remains the most lethal cancer among all cancers, it had 2.5 million cases worldwide in 2022, with more than 1.8 million deaths just in that year. The datasets used in this research are public and hence no restrictions are exercised on the utilization of the data. Validation has thus far been carried out comparing EfficientNetB5 against another pre-trained models, InceptionV3, ResNet50 and Vision Transformer. In particular, all pre-trained models were subjected to extensive experimentation to demonstrate the need for achieving a combination of explain ability plus technical accuracy ensuring trust in AI systems for a patient-centered healthcare, and thereby pushing naturally to become a new reality with this medical season