Estimation of the curved body length of tuna larvae from microscope images using a zero-shot model and image processing techniques

Abstract

Fixed larval specimens often shrink and curve, making length measurement labor-intensive. Although recent studies have demonstrated efficient fish-length estimation from images using deep learning, methods for estimating curved length remain limited. Furthermore, although deep learning is a powerful method for object detection in images, an essential step for length measurement, it requires preparing large amounts of training data, which can hinder practical implementation. In this study, we used a zero-shot model that requires no training to detect fish in an image. The curved length was then estimated using an image-processing approach that combines image thinning with Bezier curve approximation, and its accuracy was evaluated. We analyzed 1,040 larvae from five tuna species captured in stereomicroscope images. Manual measurements (notochord length or standard length; 1.5-8.5 mm) were conducted by two measurers and served as reference values. Fish regions were detected using GroundedSAM, and curved body centerlines were extracted through image thinning and approximated with Bezier curves. The curve length was used as the estimated body length. Estimation accuracy was assessed using bias and standard deviation between estimated and manual measurements. GroundedSAM detected all 1,040 fish, although there were 49 overdetections. Overdetection was caused by the double-detection of a single fish or by the misidentification of debris and light reflections as fish. Although the standard deviation of the differences between manual measurements and the image analysis-based (IAB) method was larger than the inter-measurer differences, the bias for [≤]5 mm was comparable to or smaller than the inter-measurer bias. According to the strength-of-agreement criteria for the concordance correlation coefficient, the IAB method demonstrated substantial agreement in the [≤]5.0-mm range. The IAB method accurately measured most curved tuna larvae without prior training, particularly in the [≤]5.0-mm range. Combining the IAB method with manual remeasuring can improve the efficiency of curved-length measurement tasks.