Fig 1.
Experimental design process flowchart.
The figure shows the experimental design of the study from acquisition of spectral data to machine learning training and evaluation.
Fig 2.
FNN design architecture.
Fig 3.
RNN design architecture.
Fig 4.
CNN design architecture.
Table 1.
Neural network hyperparameters for GA.
Table 2.
Hyperparameters of GA design.
Fig 5.
Median ATR-FTIR absorbance spectra of malignant (n = 53) and benign (n = 65) lung tissue samples.
The figure shows the median FTIR spectrum of malignant and benign lung tissue samples and their corresponding peaks identified via visual analysis.
Table 3.
Computation of the spectrum variables (peak positions and normalized absorbances) of malignant and benign lung samples in the fingerprint IR region (1800 cm-1 to 850 cm-1).
Fig 6.
PCA biplot showing the distribution of malignant and benign samples and the variances contributed by each biomolecule.
The red points represent the malignant samples while the blue points represent the benign samples.
Fig 7.
Average performance accuracy of NN models per generation.
The plots show the average accuracy of each NN model during the GA-based NN hyperparameter tuning process. The averaged metric shown for each generation is derived from the metric of the best individual over 50 trials. Evident in the GA plots, the FNN models were the fastest to achieve steady-state performance while the RNN model was the slowest. The RNN plot also shows a comparatively larger range of values per generation, which may suggest that the search space for RNN models of very high accuracy is relatively smaller than those of the FNNs and the CNN; hence RNN models may be the most difficult to tune. A. Average performance accuracy of FNN2-type individuals per generation. B. Average performance accuracy of FNN4-type individuals per generation. C. Average performance accuracy of FNN8-type individuals per generation. D. Average performance accuracy of CNN-type individuals per generation. E. Average performance accuracy of RNN-type individuals per generation.
Table 4.
FNN hyperparameters.
Table 5.
RNN hyperparameters.
Table 6.
CNN hyperparameters.
Table 7.
Mean and standard deviation of diagnostic performance of all the machine learning models.
Table 8.
Difference of average performance metric between NN models and the SVM model.
Fig 8.
Distribution of discordant samples from concordant samples via PCA.
The plot shows the diagnosis of the models per discordant samples (dark-colored points) over the distribution of concordant malignant samples (light red) and concordant benign (light blue) samples. The diagnoses of the models for all discordant samples were consistent with the original diagnoses of the study sites.
Fig 9.
Prediction probability of NN models per discordant samples.
The figures show the prediction probability of each NN model for the discordant benign (n = 13) and discordant malignant (n = 11) samples. The discordant samples were grouped according to the diagnosis by the pathologist of their respective study sites. All NN models show a median prediction score that is above the 0.5 (50%) mark, meaning that all the models had the same prediction as that of the diagnosis of the pathologist. A. Prediction probability of FNN2 models per discordant samples. B. Prediction probability of FNN4 models per discordant samples. C. Prediction probability of FNN8 models per discordant samples. D. Prediction probability of CNN models per discordant samples. E. Prediction probability of RNN models per discordant samples.