Ensembles of Radial Basis Function Networks for Spectroscopic Detection of Cervical Pre-Cancer

Kagan Tumer, Nirmala Ramanujam, Joydeep Ghosh, Rebecca R. Richards-Kortum

Abstract

Cervical cancer remains a significant global health burden, primarily due to the limitations of current screening methods like Pap smears and colposcopy. This study presents a "one-step" neural network ensemble approach using Radial Basis Function (RBF) networks to classify in vivo fluorescence spectra of cervical tissue. By pooling the outputs of multiple RBF networks, the researchers achieved 91% sensitivity in detecting Squamous Intraepithelial Lesions (SIL), significantly outperforming traditional clinical standards and offering a robust, automated diagnostic tool for near real-time clinical use.

Introduction

Early detection of Squamous Intraepithelial Lesions (SIL) is critical for preventing cervical cancer. However, current "gold standards"—Pap smears and colposcopy—suffer from high labor intensity, human error, and low specificity. In vivo fluorescence spectroscopy offers a non-invasive alternative by probing the biochemical state of tissue. The challenge lies in interpreting complex spectral data to differentiate healthy tissue from pre-cancerous tissue without relying on subjective human judgment or rigid linear statistical models.

Dataset and Features

The study utilized a dataset comprising 361 cervical sites from 95 patients. Spectra were collected at three excitation wavelengths (337, 380, and 460 nm). The sites were categorized via histopathology into five groups, which were then mapped to target classifications of "SIL" or "non-SIL" for binary classification tasks.

Table 1: Distribution of tissue samples across histopathological categories and target classifications for training and test sets.

Methodology

The researchers employed Radial Basis Function (RBF) networks, which use local kernels to handle varying class sizes more effectively than standard Multi-Layered Perceptrons (MLP). To enhance reliability, an ensemble of 20 RBF networks was used, with outputs combined via averaging or median pooling to reduce variance. This ensemble approach mitigates the risk of a single network being "unlucky" due to initial random weights.

A key innovation was the transition from a complex "two-step" diagnostic pipeline to a "one-step" ensemble classifier. Previous statistical methods required filtering out different healthy tissue types (Normal Squamous vs. Normal Columnar) before identifying SIL. The RBF ensemble successfully handles this complexity in a single step.

Figure 1: The traditional two-step diagnostic algorithm which the proposed RBF ensemble simplifies into a single step.

Results and Discussion

The RBF ensemble achieved a sensitivity of 91%, a marked improvement over the 62% sensitivity typical of human-read Pap smears. Specificity remained competitive at 65-67%, which is comparable to Pap smears (~68%) and significantly better than colposcopy (~48%).

Furthermore, the ensemble approach demonstrated superior stability and reliability. The study also showed that dimensionality reduction—reducing 160 spectral features to just 13—maintained high performance. This finding is crucial for the development of cost-effective, portable diagnostic hardware that can be used in point-of-care settings.

Conclusion

Ensembles of RBF networks provide a powerful framework for the automated detection of cervical pre-cancer. By simplifying the diagnostic pipeline and increasing sensitivity, this method addresses the primary weaknesses of current screening protocols, paving the way for real-time, objective clinical applications.

Summary updated

Abstract

The mortality related to cervical cancer can be substantially reduced through early detection and treatment. However, current detection techniques, such as Pap smear and colposcopy, fail to achieve a concurrently high sensitivity and specificity. In vivo fluorescence spectroscopy is a technique which quickly, non-invasively and quantitatively probes the biochemical and morphological changes that occur in pre-cancerous tissue. A multivariate statistical algorithm was used to extract clinically useful information from tissue spectra acquired from 361 cervical sites from 95 patients at 337, 380 and 460 nm excitation wavelengths. The multivariate statistical analysis was also employed to reduce the number of fluorescence excitation-emission wavelength pairs required to discriminate healthy tissue samples from pre-cancerous tissue samples. The use of connectionist methods such as multi layered perceptrons, radial basis function networks, and ensembles of such networks was investigated. RBF ensemble algorithms based on fluorescence spectra potentially provide automated, and near real-time implementation of pre-cancer detection in the hands of non-experts. The results are more reliable, direct and accurate than those achieved by either human experts or multivariate statistical algorithms.
BibTeX
@article{tumer1998ensembles,
  author = {Kagan Tumer and Nirmala Ramanujam and Joydeep Ghosh and Rebecca R. Richards-Kortum},
  title = {{Ensembles of Radial Basis Function Networks for Spectroscopic Detection of Cervical Pre-Cancer}},
  journal = {IEEE Trans. Biomedical Engineering},
  year = {1998},
  doi = {10.1109/10.704864},
}
Figure