Application of Neural Networks in Breast Cancer Prediction: From Data to Clinical Decision-Making

This article introduces an open-source project using neural networks for breast cancer prediction, aiming to analyze medical data through machine learning techniques and assist in early cancer screening and diagnostic decision-making. The project covers the complete process including data processing, model construction, training and evaluation, and discusses its technical implementation and clinical application value.

Currently, breast cancer screening relies on technologies such as mammography and ultrasound, but there are issues like diagnostic complexity (high false positive/negative rates due to overlapping benign and malignant images, large differences among readers) and underutilization of data (multi-dimensional patient data not effectively mined).

The project uses a public breast cancer dataset containing cytological features from fine-needle aspiration biopsies of breast masses. Features include 10 indicators of nuclear morphology (radius, texture, perimeter, area, smoothness, compactness, concavity, concave points, symmetry, fractal dimension), each with mean, standard deviation, and worst value, forming a total of 30-dimensional feature vectors; the target variable is binary (Malignant M/Benign B).

The project adopts a Multi-Layer Perceptron (MLP): the input layer receives 30-dimensional features, hidden layers use ReLU activation (simple computation, no gradient saturation, accelerates convergence), and the output layer uses Sigmoid to output malignant probability; the loss function is binary cross-entropy, and the optimizer is Adam. Data preprocessing includes missing value handling (mean filling or deletion for small amounts), Z-score normalization (making features follow a distribution with mean 0 and standard deviation 1), and stratified sampling to split training/test sets (ensuring consistent sample proportions).

The training process monitors training loss and validation loss curves to avoid overfitting; evaluation uses confusion matrix and multiple metrics: accuracy (overall correct proportion), precision (proportion of actual malignancies among predicted malignancies), recall (proportion of correctly identified malignancies among actual malignancies), F1 score (harmonic mean of precision and recall), and AUC-ROC (comprehensive measure of model discrimination ability). Recall is more valued in medical scenarios (cost of missed diagnosis is higher than misdiagnosis).

Model advantages: objectivity (eliminates subjective bias), consistency (same input leads to same output), scalability (integrates into automated systems), and continuous learning ability (optimizes with data accumulation). Limitations: depends on training data distribution (may not apply to data from different populations/devices), decision black box is hard to explain, and it is only an auxiliary tool (cannot replace doctors' independent diagnosis).

Application scenarios include: screening assistance (marking high-risk cases to help doctors focus on difficult cases), decision support (providing malignant probability references for borderline cases), and training and education (helping residents understand the correlation between features and malignancy).

Machine learning is developing rapidly in the field of medical diagnosis, and this project demonstrates the potential of neural networks in processing medical data. In the future, we need to accumulate high-quality annotated data and improve model interpretability; artificial intelligence is expected to play a greater role, but final health decisions still require doctors' professional judgment and humanistic care.