Computer-aided classification of breast cancer nuclei

Area of Science:

• Oncology
• Biomedical Engineering
• Computer Science

Background:

• Breast cancer is a leading malignancy in women, with prognostic factors like steroid receptors crucial for treatment decisions.
• Manual assessment of steroid receptors in biopsy slides is subjective, leading to interobserver and intraobserver variability.
• Existing computerized systems lack standardization and their outputs are not easily interpretable by medical experts.

Purpose of the Study:

• To introduce a novel computer-aided system for classifying breast cancer nuclei using neural networks.
• To develop a nuclear classification module that translates low-level image features into a standardized grading scheme.
• To improve the accuracy and consistency of breast cancer grading compared to manual methods.

Main Methods:

• A feedforward neural network was trained in a supervised manner to classify nuclear feature vectors.
• The system utilizes six input features: optical density, chromaticity indices, texture measures, and nuclei density.
• The output is a membership label in a zero to four grading scheme for each detected nucleus, based on 3015 nuclei from 28 images.

Main Results:

• The neural network achieved 72% accuracy in classifying breast cancer nuclei.
• A Sammon plot visualization indicated the complexity of the six-dimensional input feature space.
• The system successfully translated low-level measurements into a grading scheme comprehensible to medical experts.

Conclusions:

• The developed computer-aided system offers a standardized approach to grading breast cancer nuclei.
• This system aids in overcoming the limitations of manual assessment and enhances the interpretability of automated analyses.
• The nuclear classification module contributes to more accurate and consistent breast cancer diagnosis and prognosis.