White learning methodology: A case study of cancer-related disease factors analysis in real-time PACS environment

Background and objective: Bayesian network is a probabilistic model of which the prediction accuracy may not be one of the highest in the machine learning family. Deep learning (DL) on the other hand possess of higher predictive power than many other models. How reliable the result is, how it is deduced, how interpretable the prediction by DL mean to users, remain obscure. DL functions like a black box. As a result, many medical practitioners are reductant to use deep learning as the only tool for critical machine learning application, such as aiding tool for cancer diagnosis.

Methods: In this paper, a framework of white learning is being proposed which takes advantages of both black box learning and white box learning. Usually, black box learning will give a high standard of accuracy and white box learning will provide an explainable direct acyclic graph. According to our design, there are 3 stages of White Learning, loosely coupled WL, semi coupled WL and tightly coupled WL based on degree of fusion of the white box learning and black box learning. In our design, a case of loosely coupled WL is tested on breast cancer dataset. This approach uses deep learning and an incremental version of Naïve Bayes network. White learning is largely defied as a systemic fusion of machine learning models which result in an explainable Bayes network which could find out the hidden relations between features and class and deep learning which would give a higher accuracy of prediction than other algorithms. We designed a series of experiments for this loosely coupled WL model.

Results: The simulation results show that using WL compared to standard black-box deep learning, the levels of accuracy and kappa statistics could be enhanced up to 50%. The performance of WL seems more stable too in extreme conditions such as noise and high dimensional data. The relations by Bayesian network of WL are more concise and stronger in affinity too.

Conclusion: The experiments results deliver positive signals that WL is possible to output both high classification accuracy and explainable relations graph between features and class.

Keywords: Bayesian network; Data mining methodology; Deep learning; Radiological data analysis.