Enhanced segmentation of gastrointestinal polyps from capsule endoscopy images with artifacts using ensemble learning

Figure 1 Examples of artifact-infested endoscopy images used in our current study. A: Presence of cloudy liquid and specular reflections; B: Presence of bubbles and low contrast between lesion tissue and normal tissues; C: Presence of fecal residue; D: Presence of bubbles and specular reflections.

Figure 2 Semantic segmentation problem in the field of computer vision.

Figure 3 Learning rate for each epoch in the snapshot ensemble learning, with γ and α₀ set for 0. 3, M for 10, and T for 75.

Figure 4 Validation loss in training.

Figure 5 Train-test workflow of the proposed method.

Figure 6 Illustration of intersection over union and dice metrics. IoU: Intersection over union; FN: False negative; TP: True positive; FP: False positive.

Figure 7 Comparison between the ensemble and single models on artifact-affected images. Origin represents the images taken by the endoscope. Ground truth represents the anomalous locations marked by experts based on the original images. Val. 1^st best to Val. 5^th best represents the results of the five best-performing single models evaluated on the validation set. 5-ensemble represents the results of the ensemble model based on these single models.

Figure 8 Comparison between the ensemble and single models on standard (clear) images. Ground truth represents the anomalous locations marked by experts based on the original images. Val. 1^st best to Val. 5^th best represents the results of the five best-performing single models evaluated on the validation set. 5-ensemble represents the results of the ensemble model based on these single models.