SpotLab/filarias_species_detection

This model is an object detection model trained with tensorflow object detection API, published with the paper Edge Artificial Intelligence for real-time automatic quantification of filariasis in mobile microscopy

• —Model description:
• —Developed by: Spotlab
• —Model type: SSD mobilenet v2
• —Model input: image resized to 640 and normalized to with mean=127.5 and std = 127.5.
• —Classes: Loa loa, Mansonella perstans, Wuchereria bancrofti, Brugia malayi
• —Datasets:
• —Training set: 1203 field of view images (400 magnification) from 85 independent samples with 906 L. loa, 378 M. perstans, 35 W. bancrofti, and 58 B. malayi parasites.
• —Validation set: 311 field of view images (100 magnification) from 30 independent samples with 138 L. loa, 102 M. perstans, 29 W. bancrofti, and 5 B. malayi parasites.
• —Test set: 624 field of view images (100 magnification) from 18 independent samples with with 658 L. loa, 15 M. perstans, 21 W. bancrofti, and 23 B. malayi parasites.
• —Performance:
• —On validation set: the species differentiation algorithm achieved a weighted precision of 84.08%, recall of 95.33%, and an F1 score of 94.70%. Breaking down the results per class, the precision rates were 94.85% for L. loa, 97.03% for M. perstans, 94.00% for W. bancrofti, and 66.67% for B. malayi. The corresponding recall rates were 93.48%, 96.08%, 97.92%, and 92.31% respectively.
• —On test set: overall precision of 95.46%, recall of 97.81%, and F1-score of 96.62%. The per-class precision values were determined as 98.80% for L. loa, 60.00% for M. perstans, 100.00% for W. bancrofti, and 58.97% for B. malayi. The corresponding recall rates were calculated as 98.50%, 100.00%, 76.00%, and 100.00%, respectively.

Example predictions:

You can create your own android app to run this model following this tutorial: (TensorFlow Lite Object Detection Android Demo )[https://github.com/tensorflow/examples/tree/master/lite/examples/object_detection/android]