No abstract

Purpose/Objective(s): To correlate parameters derived from apparent diffusion coefficient (ADC) volumes associated with the changes in specified abnormal white matter (WM) to overall survival (OS) following therapies in glioblastoma (GBM). The unsupervised machine learning was designed to reduce uncertainties on a relatively small size data for a prediction model. Materials/Methods: The study employed 41 GBM patients treated with radiotherapy (60 Gy) followed by Temozolomide. The ADC ranges for normal WM and gray matter (GM) were established from a pool of ADC voxels over 66 individual brains. The predictors were the averaged values of pre/post treated (TX) over the masked volumes, where (abnormal) pre-TX WMs were in normal GM ADC range, then applied the mask to the post-TX volume. The learning process has 2 main parts: (1) the probability of mapping from predictors to OS, (2) data resampling extracted the features of the mapping in parametric space with different complex levels of models. The number of coefficients in linear models can be from 3 to 7 (or more) for higher complexity. The probability was based on Bayesian with likelihood probability from institutional data and the prior information from published OS reports. The data were resampled with total number and location of datapoints being altered with bandwidth of kernel density estimation, cluster analysis in parametric space. Results: The characterized ADC volumes(1) provide more objective parameters for machine learning than manually contoured clinical tumor volume (CTV), (2) are within the sampled ADC volume based on FLAIR & T1-post highlighted regions, where the ADC histogram illustrates a unique pattern of pre-/post-TX dynamics; i.e. "WM looking like GM before treatment". Without resampling in the learning process, the predicted OS had 0.57 of correlation coefficient (CC) with the data fitted with the simpler model. The model underestimated survival at late time points (>450 days) and overestimated survival in early time points. With resampling, the model CC improved (CC w 0.65) secondary to outlier suppression at later time points (survival: 200w450 days). The use of more sophisticated models did not improve the model fit. Conclusion: The subtle changes detected by ADC parameters could not be accurately quantified by T1, T2 or FLAIR suggesting that ADC offers additional information useful for prognostication. Despite high uncertainties presented in the small dataset, the correlation was non-trivial, and it can also be enhanced further with the learning process to feature the given data and a priori. Because the WM scaled as GM by ADC indicates WM abnormities, the relation of "quantitative abnormity" with OS could be pathologically meaningful. The algorithm appears to represent an optimum on the likely feature that data given as more complex models were limited by the small data size.Purpose/Objective(s): Stereotactic body radiation therapy (SBRT) has become an acceptable treatment modality for medically inoperable stage I non-small cell lung cance...

No abstract

No abstract