Mining of Training Samples for Multiple Learning Machines in Computer-Aided Detection of Lesions in CT Images

Optimal selection of training samples is very difficult when multiple learning machines are used in classification. We investigated an approach to mining of training samples for multiple learning machines in computer-aided detection of lesions. Our approach starts from "weakness" analysis of a seed machine-learning (ML) model trained for a given task. The weakness is analyzed in the receiver-operating-characteristic (ROC) space in classification. The most to least "difficult" samples for the seed model are "mined" by dividing samples into N groups by the ROC scores. N ML models are trained with the mined N groups of training samples in an ensemble manner. We tested our approach in classification between 25 lesions and 489 non-lesions. Our ML ensemble trained with the mined samples achieved a performance higher than did an ML ensemble with manually selected training samples.