Explainable AI for Medical Diagnosis: Balancing Performance and Interpretability in Deep Learning Models

Deep learning models have demonstrated remarkable performance in medical diagnosis tasks, often matching or exceeding expert clinicians. However, their widespread adoption in clinical practice remains limited, primarily due to their “black-box” nature, which lacks transparency in decision-making processes. This paper introduces MEDXAI, a novel framework that balances diagnostic performance with interpretability for deep learning models in medical applications. Our approach integrates three complementary explanation methods: (۱) attention-guided feature attribution that highlights influential image regions, (۲) concept-based explanations that map network activations to human-understandable medical concepts, and (۳) counterfactual explanations that demonstrate how alterations to input data would affect diagnostic decisions. We implement MEDXAI on three medical imaging tasks: chest X-ray classification, skin lesion diagnosis, and brain tumor segmentation. Quantitative evaluation demonstrates that our models maintain diagnostic accuracy comparable to state-of-the-art black-box models (average decrease of only ۱.۸%) while providing explanations that significantly improve clinician understanding and trust. In a user study with ۴۲ medical professionals, MEDXAI explanations improved diagnostic decision-making accuracy by ۱۶% and reduced decision time by ۲۹% compared to non-explainable models. Furthermore, our framework provides a tunable transparency parameter that allows clinicians to adjust the trade-off between model performance and interpretability based on specific clinical needs. These results demonstrate that explainable AI approaches can be successfully integrated into high-performance diagnostic models, potentially accelerating the adoption of AI systems in clinical practice while ensuring responsible and trustworthy deployment.

کلیدواژه ها:

Explainable AI ، Medical Diagnosis ، Deep Learning ، Interpretability ، Healthcare AI ، Medical Imaging ، Clinical Decision Support ، Computer-Aided Diagnosis