ROC-Score-Based Ensemble Training for Multiple Deep Learning Modules in Classification Between Polyps and Non-Polyps in CT Colonography

We developed an automatic ensemble training method for fusing massive-training artificial neural network (MTANN) deep-learning modules in classification between polyps and non-polyps in CT colonography. We started from an initial MTANN module that had been trained with an initial set of polyps and non-polyps. We applied the trained initial module to polyps and non-polyps to analyze the weakness of the initial module. We arranged the output scores of the initial module to form a score scale in receiver-operating-characteristic (ROC) space, representing the “degree of difficulty” in distinction between polyps and non-polyps by the initial module. Based on the score-space, several sets of training polyps and non-polyps with different degrees of difficulties were determined. We trained several MTANN modules with the several sets of training samples so that each module became an expert at a certain level of difficulty. We then combined expert modules with a mixing module to form a “mixture of expert” MTANNs. Our database consisted of CT colonography datasets acquired from 100 patients, including 26 polyps. The mixture of expert MTANNs with the ensemble training method distinguished all polyps correctly from more than 50% of the non-polyps. We compared the effectiveness of the ensemble training with that of training with manually selected cases. The performance of the mixture of expert MTANNs with ensemble training method was superior to that of the “reference-standard” MTANNs trained with manually selected cases, which could reduce the cost of the manual selection.