Figure 6 | Scientific Reports

Figure 6

From: PhagoStat a scalable and interpretable end to end framework for efficient quantification of cell phagocytosis in neurodegenerative disease studies

Figure 6

Automated DL model optimization using feature maps: balancing feature map quality (MSE) and execution time (sec) in Unet Models We illustrate the evaluation of various quantitative metrics including (from the left) the Mean Squared Error (MSE) of feature Map signal quality compared to 30M Unet model (reference model), execution time (seconds), and GPU memory utilization (bytes). A composite score is calculated using the formula: \(\alpha \times \text {time} + \beta \times \text {memory} + \gamma \times \text {MSE}\), all metrics are min-max normalized (to make them range between 0 and 1). If a lower value indicates better performance, the normalized metric was adjusted to 1 minus its value. With \(\alpha = 0.5\), \(\beta = 0\), and \(\gamma = 0.5\). The red point indicates the best trade-off achieved between execution time and feature map signal quality, presenting the ideal parameters for the Unet model.

Back to article page