Colony Binary Classification Based on Persistent Homology Feature Extraction and Improved EfficientNet

Classifying newly formed colonies is instrumental in uncovering sources of infection and enabling precision medicine, holding significant clinical value. However, due to the ambiguous features of early-stage colony images in culture dishes, conventional computer vision (CV) classification algorithms are often ineffective. To achieve accurate and efficient colony classification, this paper proposes a high-precision method based on Persistent Homology (PH) and an improved EfficientNet. Specifically, (1) a PH feature extraction algorithm is applied to Candida albicans (CA) and Staphylococcus epidermidis (SE) colonies cultured for 18 h in Petri dishes to capture their topological information. (2) The Mobile Inverted Bottleneck Convolution (MBConv) module in EfficientNet is modified, enhancing the attention mechanism to better handle local small targets. (3) A novel self-attention mechanism named the Spatial and Contextual Transformer (SCoT), which is introduced to process information at multiple scales, increasing the resolution in orthogonal directions of the image and the aggregation capability of feature maps. The proposed approach achieves a high accuracy of 98.64%, a 10.29% improvement over the original classification model. The research findings indicate that this method can effectively classify colonies with high efficiency.