Efficient computer-aided cervical cancer detection can improve both the accuracy and the productivity of cytotechnologists and pathologists. Nuclear segmentation is essential to automated screening, and is still a challenge. We propose and demonstrate a novel approach to improving segmentation performance by multispectral imaging followed by unsupervised nuclear segmentation relying on selecting a useful subset of spectral or derived image features. In the absence of prior knowledge, feature selection can be negatively affected by the bias, present in most unsupervised segmentation, to erroneously segment out small objects, yielding ill-balanced class samples. To address this issue, we first introduce a new measurement, Criterion Vector (CV), measuring the distances between the segmentation result and the original data. This efficiently reduces the bias generated by feature selection. Second, we apply a novel recursive feature selection scheme, to generate a new feature subset based on the corresponding CV, ensuring that the correct part of the initial segmentation results is used to obtain better feature subsets. We studied the speed and accuracy of our two-step algorithm in analyzing a number of multispectral Pap smear image sets. The results show high accuracy of segmentation, as well as great reduction of spectral redundancy. The nuclear segmentation accuracy can reach over 90%, by selecting as few as 4 distinct spectra out of 30.