Transferability of a convolutional neural network to characterise colorectal polyps

Introduction There is a lack of studies evaluating the transferability of polyp characterisation artificial intelligence systems to different populations from the institution where the training data was collected. We aimed to develop a convolutional neural network (CNN) to characterise colorectal polyps as adenoma and non-adenoma using data from two institutions (UK, Czech Republic) and to assess its transferability to a new patient population (Spain). Methodology High-quality and moderate-quality images in narrow-band imaging (NBI) and NBI-Near Focus were annotated with bounding boxes around polyps and labelled with histopathology. These were referenced as the gold standard. We developed a ResNet-101 CNN using 16,832 frames from 229 polyp videos (London, UK) and 451 still images from 266 polyps (Hradec Kralove, Czech Republic). We assessed the CNN against two internal and one external dataset (Table 1); (1) Test-set I (London), consisted of 157 polyp videos (111 diminutive), including 14,320 video frames (Olympus 260 + 290) (2) Test-set II (Hradec Kralove) consisted of 250 polyps (125 diminutive), including 487 still frames (Olympus 180 + 190) (3) Test-set III (Basque), the publicly accessible PICCOLO dataset, consisted of 53 polyps, including 855 frames (Olympus 190). Results On the per-frame analysis, the sensitivity for adenoma characterisation was 92% in test-set I and 90% in test-set II, 89% and 85% specificity, and 96% and 93% area under a curve (AUC). For the external test-set III, the CNN characterised adenomas with 86% sensitivity, 98% specificity and 99% AUC. Conclusion A CNN trained using data from two nations transferred well to an external third patient population.