The accumulation of excess fat in fish might impair the health of fish in aquaculture.
This paper introduces an online sequential extreme learning machine (OS-ELM) into
region-of-interest (ROI) detection of adipose tissues in fish digitalized by means
of magnetic resonance imaging (MRI). Three typical economic fish species, turbot (Scophthalmus
maximus L.), large yellow croaker (Pseudosciaena crocea R.) and Japanese seabass (Lateolabrax
japonicus), were selected to compose into digital physiological atlas. We manually
labelled with ITK-SNAP discriminating adipose tissue regions as standard references.
Then, single-hidden-layer feedforward neural networks (SLFNs) were established to
deduce the potential mathematical criterion for fat detection via OS-ELM for each
fish species. We further carried out classical adaptive segmentation to extract details
in fat location and distribution of adipose tissues. The quantitative correspondence
regarding adipose tissues regions, between 3D voxel representation in MRI and chemical
measurement in real fish, have been statistically investigated across each species.
The experimental results showed that our online fat detection automatically through
MRI is consistent with the standard references, and the recognition rate for three
fish species could be up to 89.13% +/- 5.32%, 91.43% +/- 6.68% and 93.08% +/- 6.57%
on average, with FAR rate 5.35%, 4.05%, 3.39% and FRRs of 5.52%, 4.52% and 3.53% respectively.
Those 3D volumes involved in fat region counting keep pace with the real weights of
adipose tissues across species, which implies we might utilize 3D voxel counting to
quantify fat accumulation in adipose tissues in a species-dependent manner. The proposed
mechanism brings comparative performances for fat detection and evaluation at a much
faster speed, which could help high-throughput insights into fat metabolism process