An improved system for lung cancer diagnosis using lung cell images

Using the same data set and techniques for an automated end-to-end lung cancer detection system, we have improved the accuracy of that system by introducing an important pre-processing stage. This system which originally consisted of Adaptive C-Means clustering for segmentation with modifications, feature extraction, and classification via Support Vector Machines, was modified to include Mean Shift segmentation as its initial stage. This addition improved the accuracy of classification from 78% to 81% when using one-hold-out cross-validation to classify Adenocarcinoma vs. normal lung cells, and up to 84% when using five-fold cross-validation. For classifying cancerous vs. non-cancerous cell images, this method also increased the cited classification from 60% to 78% using five-fold cross-validation.