The imprudent use of antibiotics has led to increased bacterial resistance, posing a serious threat to public health. Traditional visual-based antibiotic susceptibility testing (AST) methods are time-consuming and labor-intensive. Deep-learning-based bacterial detection algorithms have significantly improved the efficiency and accuracy of AST. However, existing deep-learning-based bacterial detection algorithms ignore unique challenges of AST images, such as confluent clusters, complex backgrounds, and morphologically similar noise. To address these issues, we propose a novel bacterial detection framework named the triple-stage contrastive learning framework with across-growth-stage information (TrCL-AGS). Specifically, we design a novel across-growth-stage pretraining (AGSP) task to mine across-growth-stage information from the continuous images, providing a richer and more robust feature space for subsequent stages. Instance-level contrastive learning (ILCL) is constructed to extract invariant semantic features of bacteria at a particular growth stage and exclude features similar to bacteria but do not match the specific growth stages. Besides, discriminative features between objects and backgrounds are enhanced by the category-level contrastive learning (CLCL) stage. Extensive experiments on a real-world clinical bacterial AST dataset demonstrate that our method can achieve state-of-the-art performance compared with existing object detection methods. And our framework is compatible with most detectors.

The sperm chromatin dispersion (SCD) test is widely used to assess sperm DNA fragmentation by observing the halo structure around the sperm head. However, manual halo interpretation is labor-intensive and subjective, which limits its consistency and throughput in clinical practice. To address these limitations, we propose SHIELD-Net, an image-level single-sperm halo morphology classification framework for four halo categories: large halo, medium halo, small halo, and without halo. SHIELD-Net includes two modules, i.e., a hybrid ROI extraction (HyREx) module and dual-view learning with ROI-guided gate fusion (RGF). The HyREx module converts the traditional direct classification pipeline into a segmentation-classification pipeline. It uses the cell segmentation foundation model, i.e., Cellpose, to segment halo-centered regions of interest (ROIs) and proposes a fallback mechanism to ensure segmentation effectiveness. In this way, HyREx can construct a head–halo-centered ROI and reduce interference from tail and background regions. Based on the original image and the extracted ROI, SHIELD-Net performs dual-view ConvNeXt encoding. RGF further uses the ROI representation to generate a spatial gate that recalibrates the original-view feature map toward halo-morphology-related responses. Experiments were conducted on a public expert-annotated sperm microscopy dataset containing 403 individual sperm halo images using stratified five-fold cross-validation. SHIELD-Net achieved a Macro F1 of 0.9290 ± 0.0161. Compared with the original-image ConvNeXt-Small baseline, SHIELD-Net improved Macro F1 by 0.1755. These results indicate that combining head–halo-centered ROI construction with ROI-guided feature recalibration improves image-level sperm halo morphology classification.