Bar plots of leave-one-out cross-validation error rates for k -NNs in Table 5. The classifiers were estimated from five data sets (Table 4) either without normalization (N ONRM) or normalized using twenty-three normalization techniques that remove spatial- and/or intensity effect to varying degrees (Tables 1 and 2). In each plot, the normalization methods are arranged in the following order: (A) Methods that remove no dye bias (GMEDIAN), or a single dye bias (SL LOESS, SL FILTERW3, SL FILTERW7, IG LOESS, IL LOESS, IST SPLINE). (B) Methods that remove two dye biases (IG SG LOESS, IG LOESS-SL LOESS, IL LOESS-SL LOESS, IG LOESS-SL FILTERW3, IG LOESS-SL FILTERW7, IST SPLINE-SL LOESS, IST SPLINE-SL FILTERW3, IST SPLINE-SL FILTERW7). (C) Qspline-related methods (Q SPLINEG, QS PLINER, Q SPLINEG-SL LOESS, Q SPLINEG-SL FILTERW3, Q SPLINEG-SL FILTERW7, QS PLINER-SL LOESS, QS PLINER-SL FILTERW3, QS PLINER-SL FILTERW7).