Extended performance analysis of deep-learning algorithms for mice vocalization segmentation.

Area of Science:

• Animal communication research
• Neuroscience and neuropharmacology
• Ethological studies

Background:

• Ultrasonic vocalizations (USVs) analysis is vital for studying animal communication, particularly in mice.
• Accurate USV segmentation is a critical prerequisite for reliable behavioral and neuroscience investigations.
• Current automated systems for USV detection and classification require robust segmentation methods.

Purpose of the Study:

• To evaluate the performance of three supervised deep learning methods for automated USV segmentation.
• To compare the efficacy of Auto-Encoder (AE), U-NET, and Recurrent Neural Network (RNN) for USV detection.
• To establish a benchmark for future automated USV analysis systems.

Main Methods:

• Developed and trained three supervised deep learning models: AE, UNET, and RNN.
• Input to models: Spectrograms of recorded audio tracks.
• Output: Detected regions of USV calls. Ground-truth dataset created manually.

Main Results:

• All three models achieved precision and recall scores exceeding [Formula: see text].
• UNET and AE models surpassed [Formula: see text] and outperformed state-of-the-art methods.
• UNET demonstrated superior performance on an external dataset, confirming its robustness.

Conclusions:

• Supervised deep learning methods, especially UNET and AE, are highly effective for automated USV segmentation.
• These models provide a valuable benchmark for advancing automated analysis in animal communication research.
• The findings support the use of advanced deep learning for precise USV detection in neuroscience and ethology.