Automated detection and classification of intestinal protozoan cysts using a nuclei structure-based deep learning...

Intestinal parasitic infections (IPIs) remain a critical global public health concern, particularly for epidemiological surveillance in resource-limited tropical regions. Standard laboratory diagnostics heavily rely on the manual microscopic examination of stool samples, a process that is notoriously time-consuming, labor-intensive, and highly dependent on the scarce availability of experienced parasitologists. To alleviate medical workloads and enhance community-wide screening efficiency, reliable, automated diagnostic tools are urgently needed. From an epidemiological control perspective, such tools should not only improve image-level classification but also support scalable screening, case triage, and prioritization of specimens requiring confirmatory testing in routine surveillance programs. However, most existing frameworks operate as end-to-end "black boxes" that rely solely on pixel-level data, significantly lacking the clinical explainability and morphological reasoning required for high-stakes medical decisions. To address this critical limitation, this study introduces an innovative hybrid framework that integrates deep learning algorithms with expert morphological domain knowledge to classify seven commonly encountered intestinal protozoan cysts (Entamoeba histolytica (E. histolytica), Giardia lamblia, Entamoeba coli, Endolimax nana, Entamoeba hartmanni (E. hartmanni), Iodamoeba bütschlii, and Blastocystis sp.). Utilizing a comprehensive dataset of 3143 microscopic images, the system first deploys a Mask R-CNN model to accurately detect and segment intracystic nuclei. Subsequently, a rule-based decision engine classifies the cysts by extracting key morphological attributes, including cyst shape, size boundaries, and exact nuclear configurations. The proposed framework demonstrated high diagnostic efficiency, achieving an exceptional processing speed of 0.04 s per image, making it highly scalable for high-throughput population screening. The system achieved a robust overall classification accuracy of 84.8%. While conventional light microscopy often struggles to differentiate E. hartmanni from E. histolytica due to profound structural similarities, our framework successfully resolves this taxonomic ambiguity from an epidemiological perspective. When E. histolytica and E. hartmanni were grouped into a single complex category to address this intrinsic limitation of microscopy-based methods, the overall system accuracy improved to 87.8%. This knowledge-driven AI framework provides a cost-effective, interpretable, and high-throughput screening and decision-support tool to support active disease surveillance, specimen triage, and infectious disease control programs in endemic areas.