Modified CBAM-VGG16 with Sequential SAM-CAM and Image Filtering for Automated Microscopic Fungi Classification

Fungal infections can lead to severe tissue invasion and multi-organ dysfunction, making the rapid and accurate identification of these infections critical. Traditional methods rely on culture-based and morphological examination, requiring extensive expertise. This study develops a modified attention-enhanced deep learning approach for automated microscopic fungi classification, utilizing VGG16 integrated with a decoupled Convolutional Block Attention Module (CBAM) that separates spatial and channel attention (SAM-CAM) to preserve spatial information. Preprocessing included data augmentation (rotation and flipping) to address class imbalance, as well as multiple standard image filtering techniques to enhance structural features. Three architectures— baseline VGG16, original CBAM-VGGNet, and modified CBAM-VGGNet—were systematically evaluated using transfer learning across five filtered datasets. The proposed model achieved a maximum accuracy of 94.33% with the HFE filter, outperforming the original CBAM-VGGNet (86.86%) and baseline VGG16 (78.88%). Attention map visualization and quantitative metrics (entropy and variance) confirmed that the decoupled SAM preserves spatial information, enabling more effective feature extraction. Class-wise recall improved substantially, particularly for challenging classes (from 0.418 to 0.878). These results suggest that carefully designed attention architectures, combined with appropriate preprocessing, enhance classification performance for complex microscopic specimens.