D-A GAN: A novel Dual-Attention GAN for efficient and explainable medical anomaly detection

Detecting medical anomalies is vital for enhancing diagnostic precision and streamlining clinical workflows, particularly in complex tasks such as identifying fractures and foot deformities in orthopedic imaging. This study tackles these challenges by augmenting the baseline MADGAN model with advanced Channel Attention and Spatial Attention Mechanisms. These mechanisms enable the model to dynamically prioritize relevant features and spatial regions, enhancing its ability to detect subtle anomalies across diverse imaging modalities. By focusing on both fine-grained details and global spatial dependencies, our proposed Dual Attention GAN reduces the need for additional preprocessing steps, such as region-of-interest detection, while ensuring robust and interpretable results. We trained Dual Attention GAN on three publicly available datasets: MURA, RibFrac and Pesplanus. Compared to state-of-the-art models like Res-UnitGAN and UADDGAN, Dual-Attention GAN achieves up to a 5% improvement in precision and a 6% increase in accuracy for detecting musculoskeletal anomalies, achieving competitive performance among state-of-the-art approaches in medical anomaly detection. The ablation study highlights the vital role of attention mechanisms in improving performance on anomaly detection. Moreover, Grad-CAM visualizations produce distinct and comprehensible heatmaps, facilitating the analysis and validation of the model’s predictions. Dual-Attention GAN could be a robust and explainable alternative for anomaly detection, showcasing its versatility across various imaging modalities and clinical applications and establishing a new paradigm in medical diagnostics. The code is available at https://github.com/nabinabila/Dual-Attention-MADGAN . • We present Dual Attention GAN, an innovative and advanced model specifically developed for detecting anomalies in medical imaging. • The proposed method demonstrates outstanding performance on diverse benchmark datasets. • A comprehensive ablation study highlights the significant contribution of the Dual Attention Mechanism in enhancing detection accuracy. • The model has been rigorously tested on three diverse datasets, encompassing various organs, imaging modalities, and anomaly types.