Comparison of detection performance between the proposed algorithm with conventional target detection methods and other image analysis tools
| Method | TP | FP | FN | Recall | Precision | F1 Score |
|---|---|---|---|---|---|---|
| Proposed algorithm | 6745 | 3557 | 1751 | 79.26% | 68.97% | 73.00% |
| LoG | 6064 | 8030 | 2432 | 71.02% | 54.01% | 56.46% |
| DoG | 6619 | 17358 | 1877 | 78.15% | 33.79% | 45.24% |
| DoH | 7984 | 26125 | 512 | 93.80% | 28.82% | 42.31% |
| ImageJ | 4473 | 6142 | 4023 | 52.65% | 51.42% | 47.42% |
| ilastik | 2464 | 765 | 6032 | 28.13% | 80.83% | 37.22% |
| QuPath | 3899 | 3008 | 4597 | 45.98% | 63.74% | 48.79% |
True positive (TP), false positive (FP) and false negative (FN) with averages of recall, precision, and F1 score of images using the respective methods. Our CNN-based TH+ dopaminergic neuron detection algorithm showed better performance than three conventional target detection methods: Laplacian of Gaussian (LoG), difference of Gaussian (DoG), and determinant of Hessian (DoH), and three image analysis tools: ImageJ, ilastik, and QuPath. Cellprofiler was excluded due to lack of detection performance, therefore not quantifiable. In this table, the F1 score for the proposed algorithm is not 76.51%, but 73.00%. This is because the F1 score was calculated for all images instead of the average for individual images.