Evaluation of a computational model for mycetoma-causative agents identification

Area of Science:

• Medical imaging analysis
• Computational pathology
• Machine learning in diagnostics

Background:

• Mycetoma diagnosis relies on identifying fungal (eumycetoma) or bacterial (actinomycetoma) agents.
• Current histopathological diagnosis requires skilled pathologists, scarce in endemic rural regions.
• This study addresses the need for accessible mycetoma diagnostic tools.

Purpose of the Study:

• To develop and evaluate a machine learning model for semi-automatic classification of mycetoma subtypes.
• To differentiate between eumycetoma and actinomycetoma using histopathological grain images.
• To provide an accessible diagnostic aid for mycetoma.

Main Methods:

• A computational model utilizing radiomics and partial least squares was developed.
• The model was trained and validated on 890 grains from 168 Sudanese mycetoma patients.
• The dataset included 94 eumycetoma and 74 actinomycetoma cases.

Main Results:

• The machine learning model achieved 91.89% accuracy in identifying causative agents.
• The model demonstrated robustness to minor segmentation errors and acquisition protocol variations.
• Homogeneity of mycetoma grain textures emerged as the most discriminative radiomic feature.

Conclusions:

• The computational approach offers a valuable tool for mycetoma diagnosis in resource-limited rural settings.
• This method can serve as a supplementary diagnostic aid for expert pathologists.
• Accurate identification facilitates the implementation of appropriate therapeutic strategies for mycetoma.