Accurate Bayesian Data Classification without Hyperparameter Cross-validation

Improved resection margins in breast-conserving surgery using Terahertz Pulsed imaging data

“…A multivariate Bayesian classification algorithm [7] was applied to the TPI waveform data to discriminate benign from malignant breast tissue. The classifier used supervised learning techniques with the class labels derived from the histopathologically defined tissue content.…”

“…The first is model selection. We use the framework outlined in [7]. The second stage is to train our chosen model on the training data.…”

“…This is the case with our data where n = 257 and d = 301. This is somewhat mitigated by using a fully Bayesian classification approach [7]. Inference methods such as maximum likelihood work well when n d but are prone to overfitting when n ≤ d. The classifier used in this paper avoids taking point estimates of unknown parameters.…”

“…The original THz time domain pulses were assessed in their ability to discriminate tumour from healthy (fibrous/adipose) breast tissue using our probabilistic Bayesian classification algorithm [7].…”

“…In the current paper, we aim to improve the prediction capabilities of the classifier by utilising a novel Bayesian classifier [7] which improves tissue discrimination by distinguishing between tumour, fibrous and The large number of covariates in the dataset along with a relatively small number of tissue samples makes the prediction of the clinical outcome difficult because of overfitting problems and prohibitive computation demands. Typically Bayesian calculations rely on computational methods such as Markov Chain Monte Carlo to compute integrals.…”