Page 24 - IJAMD-1-1

International Journal of AI
for Material and Design ML in 3D bioprinting of cultivated meat

Table 4. Summary of published works on the prediction of total viable count in pork
Foods ML method Analytical Sample Details Performance References
technique size
Chopped Supervised: HSI 90 Inputs: R = 0.41, RMSEP = 1.17 from image 81
2
pork BPNN • Seven PCs from 15 texture features R = 0.79, RMSEP = 0.46 from spectra
2
• Eight PCs from 111 points of HSI spectral R = 0.83, RMSEP = 0.24 from data fusion
2
data of optimal wavelengths in the range
of 450 – 900 nm, selected using siPLS.
• 11 PCs from intermediate‑level data
fusion of HSI spectral data of optimal
wavelengths ranging from 450 – 900
nm, selected using siPLS, and 15
texture features.
BPNN Model details:
• One hidden layer, eight nodes
• Tanh activation function
Minced Supervised: FTIR 134 Inputs: i. PLS-DA: 82
pork PLS-DA • First six LVs from FTIR spectral data in 88.1% overall accuracy
(Sensory the range of 1800 – 900 cm -1 86.7% accuracy for fresh
freshness), 87.5% accuracy for semi-fresh
PLSR 88.9% accuracy for spoiled
(Microbial ii. PLSR:
counts) r = 0.88, SEP = 0.67 for TVC
r = 0.87, SEP = 0.71 for Pseudomonas spp.
r = 0.83, SEP = 0.80 for
B. thermosphacta
r = 0.87, SEP = 0.67 for LAB
Raw Supervised: HSI 50 Inputs: i. PLSR: 74
pork PLSR, SVM • 2 derivative preprocessing of HSI R = 0.94
nd
spectral data in the range of Standard error = 0.74
400 – 1000 nm ii. SVM:
R = 0.94
Standard error = 0.46
Raw Supervised: E-nose 12 Inputs: i. KNN: 83
pork KNN, LDA, • E‑nose data from 11 gas sensors 77.73% accuracy for quality classification
2
SVM, MLP, KNN model details: R = 0.81, RMSE = 0.46
LSTM, • Euclidean distance, best k value in ii. LDA:
DWT-LSTM 3, 5, 7, 9, 11 76.82% accuracy for quality classification
SVM model details: R = 0.83, RMSE = 0.42
2
• Radial basis function kernel iii. SVM:
MLP model details: 84.88% accuracy for quality classification
2
• One hidden layer R = 0.91, RMSE = 0.37
• Tanh activation function iv. MLP:
LSTM model details: 56.19% accuracy for quality classification
• Three sigmoid layers, one tanh layer R = 0.41, RMSE = 10.64
2
v. LSTM:
85.14% accuracy for quality classification
R = 0.96, RMSE = 0.15
2
vi. DWT-LSTM:
94.83% accuracy for quality classification
2
R = 0.97, RMSE = 0.05
Abbreviations: BPNN: Backpropagation neural network; DWT-LSTM: Discrete wavelet transform-long short-term memory; FTIR: Fourier transform
infrared spectroscopy; HSI: Hyperspectral imaging; KNN: K-nearest neighbor; LAB: Lactic acid bacteria; LDA: Linear discriminant analysis;
LSTM: Long short-term memory; MLP: Multilayer perceptron; MSI: Multispectral imaging; PC: Principal component; PLS-DA: Partial least squares
discriminant analysis; PLSR: Partial least squares regression; ReLU: Rectified linear unit; RMSE: Root-mean squared error; RMSEP: Root-mean squared
error of prediction; SEP: Standard error of prediction; SVM: Support vector machine; TVC: Total viable count; Brochothrix thermosphacta: Brochothrix
thermosphacta.

adulteration and less so for other forms of adulteration, and to detect individual adulterants rather than the presence of
it would be of interest for further research to train models adulterants in general.

Volume 1 Issue 1 (2024) 18 https://doi.org/10.36922/ijamd.2279