Fig 1.
Experimental design process flowchart.
The figure shows the experimental design implemented in the study, from the acquisition of breast tissue samples, to the acquisition, processing and analysis of spectral data.
Table 1.
Clinical data of the patients with breast lesions*.
Fig 2.
PCA biplot showing data points of malignant and benign samples.
The red points denote malignant samples while blue points denote benign samples plotted across the two most dominant components (F1 = 90.28% and F2 = 5.21%). The vectors show the wavenumbers associated with peak absorbance, where those highlighted in green were identified as significant wavenumbers in discriminating benign from malignant samples.
Table 2.
Feedforward neural network architecture.
Fig 3.
A. Grid search surface plot of FNN2. The plot shows the grid search surface plot for optimizing the FNN2 model for learning rate and number of neurons per hidden layer. Validation accuracy peaked at a learning rate of 0.01 and 350 neurons per hidden layer. Low performance at high learning rates (blue region) may be due to divergence and high parameter oscillations, while stagnation of performance at low learning rates (green region) may be due to insufficient training time. B. Grid search surface plot of FNN4. The plot shows the grid search surface plot for optimizing the FNN4 model for learning rate and number of neurons per hidden layer. Validation accuracy peaked at a learning rate of 0.01 and 400 neurons per hidden layer. The same behavior in the low and high learning rate regions, observed in the FNN2 surface plot, is also evident here. C. Grid search surface plot of FNN8. The plot shows the grid search surface plot for optimizing the FNN8 model for learning rate and number of neurons per hidden layer. Validation accuracy peaked at a learning rate of 0.01 and 300 neurons per hidden layer. Among the FNN surface plots, the FNN8 constituted to the most unstable response in validation accuracy.
Table 3.
Diagnostic performance of the models.
Table 4.
Test of significance of neural network performance metrics relative to SVM.
Fig 4.
Median ATR-FTIR absorbance spectra of malignant (n = 88) and benign (n = 78) breast tissue samples.
The figure shows the median FTIR spectrum of malignant and benign breast tissue samples and their corresponding peaks identified via visual analysis. The plot shows almost similar absorbance among benign and malignant samples within wavenumbers associated with the amide proteins. Benign tissue samples, relative to malignant tissue samples, are shown to have increased absorbance within the region associated with lipids and nucleic acids while having decreased absorbance within the region associated with carbohydrates, glycogen, and phosphorylated proteins.
Table 5.
Comparison of the spectrum variables (peak positions and normalized absorbances) of malignant and benign breast samples in the fingerprint IR region (1800cm-1 to 850cm-1) via visual peak identification.
Fig 5.
Input response of neural networks from the designed NN-based sensitivity analysis.
The line plots show the input response of each neural network design per change of absorbance value per wavenumbers. A high per cent contribution magnitude implies a high response to a change for the particular wavenumber, hence may serve as a marker in identifying malignant samples from benign samples. As evident from the figure, the response of each network is nearly the same, which only varies slightly in the magnitude contribution.