Advancing Breast Cancer Diagnosis: Harnessing the Fast Learning Network Algorithm

By Samudrapom DamReviewed by Susha Cheriyedath, M.Sc.Jul 15 2023

In an article in press with the journal Frontiers in Oncology, researchers demonstrated the feasibility of using a fast learning network (FLN) algorithm-based breast cancer diagnosing classifier for breast cancer diagnosis.

Background

Breast cancer is one of the leading causes of death among women aged between 40 and 55 years. It is also the second most common malignancy after lung cancer worldwide. Thus, breast cancer must be detected in the early stages to increase the chance of survivability.

In the last few decades, machine learning (ML)   and data mining (DM) techniques have been effective in several fields, including voice pathology, diabetic retinopathy detection, speaker gender identification, emotion speech recognition, and coronavirus detection.

Several studies have also employed ML and DM algorithms for breast cancer diagnosis. However, most of these studies were not evaluated statistically/using sufficient assessment metrics. Recently, the FLN has gained significant attention as an effective method for data classification.

The FLN is primarily a double-parallel forward neural network (DPFNN), which is a parallel connection of a single-layer feedforward neural network (SLFN) and a multilayer feedforward neural network (FNN). The output layer of the DPFNN directly receives external information through the input layer neurons.

Additionally, the FLN hidden layer biases and input weights are produced stochastically. The weight values connecting the output layer with the input layer and the hidden layer are computed analytically using least-square methods.

In several instances, the FLN algorithm can achieve a better generalization performance with stability at high speed as it possesses fewer hidden neurons compared to other algorithms. Researchers also prefer the FLN algorithmbecause it can outperform the conventional backpropagation neural network (BPNN) and support vector machine (SVM).

The algorithm can solve issues of both binary and multi-classification tasks, eliminates overfitting, and perform as a kernel-based SVM with a neural network structure, which increases the FLN’s ability to deliver excellent learning outcomes. However, no studies have been performed using the FLN algorithm for breast cancer detection until now.

Breast cancer diagnosis using the FLN algorithm

In this study, researchers proposed using the FLN algorithm to improve breast cancer diagnosis accuracy. Two databases, including the Wisconsin Diagnostic Breast Cancer (WDBC) and Wisconsin Breast Cancer Database (WBCD), were utilized to evaluate the FLN algorithm performance.

The study employed different statistical analyses and evaluation metrics to assess the FLN algorithm performance based on two databases for the first time. The objective of the study was to develop a new FLN algorithm-based breast cancer diagnosing classifier using two separate databases. In the classification stage, the FLN algorithm was employed to diagnose whether the input sample was malignant or benign.

Several assessment measurements, including execution time, precision, receiver operating characteristic (ROC), recall, Matthews correlation coefficient (MCC), F-measure, specificity, accuracy, and G-mean, were performed to determine the performance of the breast cancer diagnosing classifier.

Additionally, the proposed classifier performance was statistically evaluated using standard deviation (STD) and root mean square error (RMSE) to verify the reliability of the results. The precision of the breast cancer classifier was assessed against the most recent studies that used the same WDBC and WBCD databases.

Significance of the study

The proposed FLN algorithm demonstrated excellent performance and successfully diagnosed breast cancer with an accuracy average of 96.88% using the WDBC database and 98.37% using the WBCD database.

Moreover, the FLN method also achieved 95.94%, 99.40%, 97.64%, 97.65%, 96.44%, and 97.85% average of precision, recall, F-measure, G-mean, MCC, and specificity, respectively, while using WBCD database, and 94.84%, 96.81%, 95.80%, 95.81%, 93.35%, and 96.96% average of precision, recall, F-measure, G-mean, MCC, and specificity, respectively, while using WDBC database.

These results displayed the reliability of the FLN algorithm as a breast cancer diagnosing classifier. The statistical analysis also confirmed the consistency of the results. The FLN output layer neurons received the external information directly through input layer neurons, which resulted in exceptional FLN performance.

Study limitations and future outlook

Some limitations of this study are as follows. Although an online breast cancer classification task was necessary, this study considered only the offline classification problem. The study also ignored the breast cancer stage classification and focused on only breast cancer detection.

Moreover, the study has not optimized the FLN algorithm in terms of random biases and input weights. The random biases and input weights cannot meet the FLN training goals/ensure that the FLN algorithm is the most effective in performing the classification. Thus, future studies must employ an optimization technique to reduce classification mistakes and produce more appropriate input weights and biases for the FLN algorithm.