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
Methods and materials
Phantom target materials: lesion mimicking serum and propanediol, were prepared in 10mm and 20mm blown thin walled glass bulbs on the end of slim capillaries. They were each immersed in the MARIA’s imaging volume, and radiofrequency measurement samples were taken at various position within the volume. Forty-eight (48) samples were collected for each target material.
Non-linear support vector machine (SVM) and linear discriminant analysis (LDA) classification were performed on channel data (178770 features), or the in-image focused frequency response (101 features) that were also subjected to further regularisation down to 4th most significant Principle Components (4 features).
Accuracy was reported by means of 25 runs of random repeated sub-sampling (Monte-Carlo simulation) with 75% and 25% of data was used for training and testing respectively for each run.
Learning (training and validation) curves were used to further predict the robustness of each technique on unseen data. Error rates were plotted as the model learns with increasing number of training examples up to 50% of available data, while the rest of the data were used to validate the model at each steps of evolution. Smaller differences in the error between training and validation curves, and convergence of the curves, indicate higher confidence in the model’s robustness towards unseen data.