CsDETC: detection and counting of small target Cryptococcus spp.

Abstract

Introduction: Cryptococcus spp. infection can lead to cryptococcal meningitis (called CM) and pulmonary cryptococcosis, and how to diagnose Cryptococcus spp. infection accurately and timely is an urgent need in clinical practice. However, the existing methods such as cerebrospinal fluid (CSF) ink staining smear microscopy and CSF Cryptococcus spp. culture only rely on manual counting to determine the number of Cryptococcus spp., resulting in low efficiency. Thus, how to identify Cryptococcus spp. in cerebrospinal fluid and achieve automated counting of Cryptococcus spp. is of great significance for helping clinical experts accurately and timely diagnose Cryptococcus spp. infections to reduce the risk of deterioration. Method: We propose a small target Cryptococcus spp. detection and counting method called CsDETC, where three important components are integrated, such as data augmentation, hypergraph computation empowered semantic collecting and scattering module called HGC-SCS, and attention-enhanced path aggregation network called AEPAN. The Cryptococcus spp. dataset has been expanded through multiple data augmentation techniques such as random cropping, horizontal flipping, and rotation before training the model. Subsequently, the Cryptococcus spp. morphological features have been enriched by data augmentation based on perspective transformation and vertical flipping in the training process, thereby improving the generalization ability. Then different morphological features can be adaptively detected by learning high-order relationships between visual features when adding HGC-SCS into the neck network. Eventually, the convolution block attention module (CBAM) is integrated into path aggregation network to generate attention maps along the channel and spatial dimensions, transmitting more detailed information contained in the shallow layer to the deep layers to enhance the perception ability of small targets. Results: The experimental results on private dataset show that CsDETC outperforms other advanced object detection models with excellent performance such as YOLOv10 and YOLO11, etc. Typically, compared to the baseline, CsDETC shows significant improvements in mAP50 (93.6% vs. 91.3%), APs (51.0% vs. 49.5%), and MAE (1.865 vs. 2.370), while only a 0.7 millisecond increase in the inference time. Discussion: CsDETC is a promising tool that has performed well in preliminary validation. After validation with larger and more diverse datasets from different medical centers in the future, CsDETC can help doctors accurately and timely identify Cryptococcus spp. and achieve automated counting of Cryptococcus spp., providing reference for treatment plans and improving the diagnostic efficiency.