Copyright ©The Author(s) 2023.
World J Gastroenterol. Jan 7, 2023; 29(1): 96-109
Published online Jan 7, 2023. doi: 10.3748/wjg.v29.i1.96
Table 1 Drawbacks of spectroscopic-related methodologies, and proposed solutions
Raman spectroscopy (random measurement points selection)(1) Not differentiating the results with cancer subtype leads to mixed results; and (2) Mixing measurements of cancer cells or tumor stroma with pancreatitis, inflammation, necrosis, colloid, etc(1) Results should be analyzed with regard to cancer subtype; (2) Raman hyperspectral mapping allows the selection of specific points of interest, whether cancer cells or tumor stroma areas; and (3) The selection should be done by an experienced pancreatic pathologist
SERS(1) Poor reproducibility, due to random and nonuniform hot spots distribution in nanostructures production[31,47,48]; (2) The distinction between methylation signals and those from adjacent nucleotides is difficult, because of their similarity[49]; and (3) Signal purity is affected by the use of surfactants and/or capping agents[50](1) PGNA as a SERS substrate; and (2) The use of a FIB to obtain a periodic matrix of holes (plasmonic nano-holes array) in a gold layer evaporated on an atomically flat non-plasmonic substrate[31]
ATR-FTIR(1) Low sensitivity and specificity; (2) Human-dependent pre-processing and analysis of spectral data; and (3) Not yet confirmed in pancreatic cancer diagnosisTo increase the sensitivity and specificity, and to limit the human intervention, convolutional neural networks are used
Multivariate data analysis(1) Results might be disturbed by the human-dependant actions and seemingly irrelevant data will be lost; and (2) Losing data lowers the sensitivity and specificity of the methodUsing of convolutional neural networks, which are fed with raw spectral data
CNN(1) Very deep neural networks are characterized by a vanishing/exploding gradient problem; and (2) Overfitting of the CNN trained on a limited amount of data(1) The use of ResNet CNN architecture, with the use of so-called “skip connections”; and (2) Proper data augmentation to prevent overfitting and sensitize the CNN to deal with various “scenarios”