Keywords:
Multicentre study, Experimental, Retrospective, Tissue characterisation, Cancer, Experimental investigations, Diagnostic procedure, Image manipulation / Reconstruction, Digital radiography, Oncology, Breast, Artificial Intelligence, Artificial Intelligence and Machine Learning
Authors:
T. Doshi1, A. W. Preece2, D. Gibbins3, L. Tsui1; 1Bristol/UK, 2Bristol, Avon/UK, 3Bristol /UK
DOI:
10.26044/ecr2020/C-11330
Results
According to Table 1, 97% accuracy advocates ‘SVM on channel data’ as the preferred solution. However learning curve data (Figure 2a) demonstrates a lack of robustness on unseen data. The error rate from training was unrealistic throughout and validation error did not reduce to training’s level, this exhibits signs of overfitting compared to other methods.
Table 1: Summary of results.
Methods |
Features |
Accuracy (Monte-Carlo simulation |
Robustness towards unseen data (learning curves) |
SVM on channel data |
178770 |
97% |
Poor |
SVM on in-image focused frequency response |
101 |
81% |
Good |
LDA on PCA regularised in-image focused frequency response |
4 |
83% |
Good |
Accuracies on in-image focused frequency response methods are similar. And their training curves (Figure 2b and 2c) show similar trends as the error rate raised to a similar level at a third of training progress, then reduced gradually towards a small increase as the models adjusted to more training samples, reducing again at the end. On studying the validation curve, results show how the models reacted to the same validation set throughout, with the SVM model showing gradual decrease in error rate while LDA (with PCA regulation) showed early optimisation then stabilised towards the end. Both plots shown signs of convergence with results matching Monte-Carlo simulation. This demonstrates good performance with confidence of robustness towards unseen data. This suggest that the focussed frequency response is optimal.