Evaluating Imaging Technologies for Non-invasive Sex Determination in Atlantic Cod

Summary

The project was motivated by the need for reliable and non-invasive methods to determine sex in farmed Atlantic cod, a capability that would support selective breeding, production planning, and improved broodstock management. Traditional approaches such as dissection and ultrasound are either destructive or difficult to apply accurately to juvenile and pre-mature fish, which limits their value in commercial settings. This project aimed to evaluate whether hyperspectral imaging, employing the Maritech Eye system, could detect biochemical or biophysical signals associated with sex that are not detectable through conventional inspection. Although hyperspectral imaging was the primary focus, other imaging based technologies including fluorescence hyperspectral imaging and deep learning using RGB images were also assessed to explore whether alternative modalities could support early and high throughput sex sorting in aquaculture. To address the objective, several experiments were carried out on fish ranging from approximately 300 g to 3 kg, spanning sizes that align with industrially relevant classes from early growth stages to more mature individuals. Hyperspectral images were acquired from cod presented with different orientation (ventral and lateral), and sex was determined through dissection to ensure accurate ground truth labels in referencing. The study applied a range of data analytical methods, including exploratory spectral analysis, supervised classification models, and a tailored exploitation of the spectral signatures from excised gonads using linear discriminant analysis (LDA). Additional imaging modalities were also explored, including fluorescence hyperspectral imaging to detect potential sex related emission patterns and deep learning models applied to RGB images to assess whether externally observable visual information (such as morphology, skin pigmentation, or textural features) could be used for sex prediction. All methods were evaluated under controlled data partitioning and consistent validation procedures to ensure robust performance assessment. The findings indicated that, although sex‑specific spectral information could potentially be present in whole fish images, it was not sufficiently strong or separable from other tissues and spatial variability to be reliably linked to sex using supervised classifiers. In contrast, spectra obtained directly from the gonads showed a clear distinction between males and females, and the LDA projection derived from these spectra demonstrated promising transferability when applied to whole fish images with well developed gonads. Fluorescence imaging did not reveal sex dependent signals, and deep learning networks trained on RGB images were unable to learn discriminative patterns for sex determination. Overall, the results indicate that while the LDA projection based on the gonad spectra shows potential for fish with developed gonads, none of the evaluated techniques currently offers a robust, noninvasive solution for early sex determination in Atlantic cod under industrial conditions.

Read publication here:

NVA : hdl.handle.net/11250/5515464