CNN-based event classification for alpha-decay events in nuclear emulsion

We developed an efficient classifier that sorts alpha-decay events from various vertex-like objects in nuclear emulsion using a convolutional neural network (CNN). Alpha-decay events in the emulsion are standard calibration sources for the relation between the track length and kinetic energy in each emulsion sheet. We trained the CNN using 15,885 images of vertex-like objects including 906 alpha-decay events and tested it using a dataset of 46,948 images including 255 alpha-decay events. By tuning the hyperparameters of the CNN, the trained models achieved an Average Precision Score of 0.740 +/- 0.009 for the test dataset. For the model obtained, a discrimination threshold of the classification can be arbitrarily adjusted according to the balance between the precision and recall. The precision and recall of the classification using previous method without a CNN were 0.081 +/- 0.006 and 0.788 +/- 0.056, respectively, for the same dataset. By contrast, the developed classifier obtained a precision of 0.547 +/- 0.025 when a similar recall value of 0.788 was set. The developed CNN method reduced the human load for further visual inspection after the classification by approximately 1/7 compared to the estimated load of the former method without a CNN.