High-throughput phenotyping by deep learning to include body shape in the breeding program of pacu (Piaractus mesopotamicus)

• Autor(a) principal: Freitas, Milena V. [UNESP]
• Data de Publicação: 2023
• Outros Autores: Lemos, Celma G. [UNESP], Ariede, Raquel B. [UNESP], Agudelo, John F.G. [UNESP], Neto, Rubens R.O. [UNESP], Borges, Carolina H.S. [UNESP], Mastrochirico-Filho, Vito A. [UNESP], Porto-Foresti, Fábio [UNESP], Iope, Rogério L. [UNESP], Batista, Fabrício M. [UNESP], Brega, José R.F. [UNESP], Hashimoto, Diogo T. [UNESP]
• Tipo de documento: Artigo
• Idioma: eng
• Título da fonte: Repositório Institucional da UNESP
• Texto Completo: http://dx.doi.org/10.1016/j.aquaculture.2022.738847; http://hdl.handle.net/11449/247644
• topic: Artificial intelligence; Genetic selection; Phenomics

Deep learning (DL) is a cutting-edge technology that enables high-throughput phenotyping in aquaculture. The routine application of DL offers new opportunities for the genetic selection of appearance traits, especially those related to body shape. The criteria currently used for the trait selection of commercial interest, such as rapid growth and weight gain, can directly influence the animal's appearance, which is a criterion for sales and profit. Different morphotypes of the pacu Piaractus mesopotamicus (elliptical and rounded) have been described previously and may represent different commercial trends. Therefore, this study aimed to 1) develop a computer vision system (CVS) through deep learning that targets the prediction of morphometric measurements and body shape (morphotypes) in pacu, 2) analyze whether morphotypes vary according to the environment, sex, and/or age, and 3) estimate genetic parameters for body shape, using the condition factor (K) and ellipticity (E) as criteria. Data from 1380 individuals corresponding to 48 full-sib families were evaluated in two distinct environments (breeding nucleus: env1; commercial fish farm: env2). The animals were evaluated based on their weight and morphometric measurements at 15 and 28 months of age (growth stage). We used the mask R-CNN model as a deep-learning algorithm, which was optimized for a ResNet architecture with only 18 layers. This resulted in a faster training period (8GB NVIDIA 2060 RTX in less than a day), which requires less computational effort. The pacu CVS was effectively developed to account for the segmentation of several fish body regions (head, body, fins, and pelvis), as corroborated by the high correlations of measurements predicted manually and automatically. We detected K and E variation at different growth stages and environments, in which fish tend to have rounded shapes in env2 and at 28 months old. The body shape heritability indicates that this trait is under moderate genetic control and should respond to selection. In conclusion, this study established an efficient CVS for pacu that is resilient to field conditions, allowing high-throughput phenotyping for the routine assessment of body shape in breeding programs for this species.