Detecting Breast Cancer Using Visual ML

Abstract

Approximately 60% of Breast cancer patients are diagnosed in advanced stages. This paper examines the automation of identification of cancerous cells using visual machine learning approach. Results are obtained using two different datasets: Wisconsin and Coimbra. In Wisconsin dataset, predictors are extracted from the digitised image of a fine needle aspirate (FNA) of a breast mass. In Coimbra dataset, predictors are extracted from the blood analysis. Ten machine learning models are compared using a visual ML tool called Orange. Particular emphasis is placed on the metric “recall”. Recall is defined as the ability to catch malignant cases out of total malignant cases present in the dataset. Highest recall of 0.982 in Wisconsin dataset is achieved using Stochastic Gradient Descent (SGD). Highest recall of 0.793 in Coimbra dataset is achieved using Gradient Boost algorithm. This tool can be deployed in hospitals where initial detection may be done using blood analysis and then for confirmation with digitised image of breast mass. In countries like India, where there is a scarcity of cancer specialists, this tool would fasten the detection process. Patients would spend more time in treatment than in diagnosis.