This application's broad, long-term objective is to diagnose cancer accurately, reducing both false-negative and false-positive diagnoses. The research proposed in this application concerns artificial neural networks (ANNs) which are important statistical tools that are used frequently in computer-aided diagnosis (CAD) methods intended to improve cancer diagnosis. The goal of this research is to provide error bars for ANN output. The significance and health-relatedness of this research is that the goal, if achieved, could represent a fundamental advance to ANNs in CAD applications, by making ANN output more reliable for subsequent computer processing and more intuitive to understand for radiologists to interpret and incorporate (the diagnostic predictions made by ANNs) in their own diagnoses. The proposed method could become the standard of practice, replacing the conventional, one-ANN approach. The hypothesis to be tested is that artificial neural network output has finite uncertainty which can be estimated and expressed in terms of confidence intervals.
The specific aims are: (1) To demonstrate qualitatively the concept of uncertainty in ANN output and the feasibility of developing multiple ANNs from a single set of training cases. (2) To develop and validate quantitative method(s) of ANN-precision estimation. (3) To apply ANN-precision estimation to a real-world CAD problem: computer classification of malignant and benign clustered microcalcifications. The research design is to use primarily computer simulations to investigate properties of the output from multiple ANNs obtained from the same training cases and to develop and validate practical method(s) for computing confidence intervals in ANN output from these multiple ANNs, then to apply the new methods to a real-world CAD task. The methods to be used include computer simulation, ROC analysis, computation of confidence intervals, bootstrapping, parametric estimation of statistical distributions, analytical analysis, and computer analysis of breast lesions in mammograms.
| Liu, Bei; Jiang, Yulei (2013) A multitarget training method for artificial neural network with application to computer-aided diagnosis. Med Phys 40:011908 |
| Zur, Richard M; Jiang, Yulei; Pesce, Lorenzo L et al. (2009) Noise injection for training artificial neural networks: a comparison with weight decay and early stopping. Med Phys 36:4810-8 |
| Jiang, Yulei (2003) Uncertainty in the output of artificial neural networks. IEEE Trans Med Imaging 22:913-21 |
