Applicability of Approximate Inverse Model Explanations (AIME) to Medical Imaging and Acceleration Using GPU

This study proposes a methodology for applying Approximate Inverse Model Explanations (AIME) to medical image analysis to enhance deep learning models’ interpretability and investigate their efficacy and computational efficiency. AIME approximates the inverse operator of the model using the Moore-Penrose pseudo-inverse matrix, facilitating the simultaneous visualization of global and local feature importance. This approach uniquely mitigates the effects of the nonlinearities and multi-collinearity inherent in complex medical image data, yielding stable explanations that conventional XAI methods cannot provide. Experiments were conducted using a DenseNet121-based chest X-ray classification model on 16,000 images randomly selected from the NIH chest X-ray dataset. In addition, GPU-based parallel computation facilitated the acceleration of large-scale matrix operations and inverse matrix computations, achieving a maximum speedup of 147.16 times in calculating local feature importance. These findings demonstrate that AIME enhances interpretability and reduces computation time for high-dimensional and complex medical image data compared to conventional methods. Moreover, this method effectively elucidates the representative feature patterns for each disease class. It evaluates the congruence between the regions of interest identified by the model and clinically relevant diagnostic indicators. The results of this study are significant not only for their application to diagnostic support systems but also as a new approach to solving the black-box nature of deep learning models. Owing to the benefits of parallel GPU computation, this opens up prospects for high-dimensional data analysis and real-time processing.