Abstract:
This study presents an explainable classifier for identifying Plasmodium species and life stages from thin smear images by integrating Convolutional Neural Networks (CNNs) with Pathogen Ontology. Using the CDC Thin Smear dataset, the model applies SegNet for pixel-wise semantic segmentation, identifying features like infected red blood cells, ring forms, and gametocytes. Ontological reasoning maps visual features to structured biological concepts, producing interpretable outputs (e.g., "Parasitized: hasCleft AND hasDots"). This approach enhances diagnostic transparency, enabling clinicians to understand the AI’s decisions. Additionally, Grad-CAM visualizations support explainability by highlighting relevant image regions, fostering trust in the system. The combination of deep learning and ontology ensures real-time, reliable malaria diagnostics, reducing human error while maintaining clinical relevance.
Description:
The detection of Plasmodium falciparum (Pf), the parasite responsible for severe cases of malaria, is a critical task in healthcare, particularly in malaria-endemic regions. Traditionally, identification of Pf is performed through manual microscopic examination of blood smear images, which is labor-intensive, time-consuming, and subject to human error. Automation of this process has gained momentum, with advances in machine learning, particularly Convolutional Neural Networks (CNNs), driving research in this domain. A recent trend within this context is the use of ontology-based frameworks combined with CNNs, which aim to enhance the explainability and transparency of model decisions, crucial for healthcare applications where interpretability is essential.
