Colonic Polyps

The training and validation datasets used for the proposed system were obtained from the PolypsSet dataset [22 (link)] and Chang Gung Memorial Hospital (Table 1). In total, 3750 low-quality images and 2500 high-quality images were selected by experienced colonoscopists from the experimental datasets. The low-quality images included blurred images and images that contained folds, fecal matter, and opaque water. High-quality images were defined as images with clear and well-distended colon lumen and with no fecal residue or opaque fluid. Each image had a resolution of 640 × 480 pixels and was a TIF file. Among the low-quality images, the number of blurred images and the number of images containing folds, fecal matter, and opaque fluid were 2500 and 1250, respectively. In addition, the number of high-quality images containing polyps and the number of those containing no polyps were 1250 and 1250, respectively. The test dataset was derived from six videos (Table 2) that had been obtained for this study from Linkou Chang Gung Memorial Hospital. Each video was in MKV format, lasted approximately 15 min, and displayed 30 frames per second. The colonoscope model was CF-H290L/I, which featured a 170° angle of view, a forward-viewing direction of view, and a depth of field of 5–100 mm. After the images were de-identified and all of the non-intestinal information was cropped from the images, the images had a resolution of 720 × 960 pixels. After the deletion of the first 3–10 min portion of each video, which showed the insertion of the colonoscope into the cecum, the remaining footage was employed for polyp detection and identification at 3 frames per second. Among the dynamic images obtained from the videos, the number of blurred images and the number of images containing folds, fecal matter, and opaque water were 8716 and 1967, respectively, and the numbers of high-quality normal images and polyp images were 399 and 50, respectively. All of the videos featured one polyp, except for Case #1, which featured two polyps (Table 3). Polyp detection was performed using a CNN model for classification, and the training dataset comprised 612 images from the CVC-ClinicDB dataset and 500 images from the PolypsSet dataset (Table 4).
Figure 2 shows the architecture of the proposed intraprocedure alert system, which provides blurred image detection, foreign body detection, and polyp detection. Blurred image detection is used to identify blurred images that have occurred due to camera shaking, to the colonoscope being withdrawn too rapidly, or to the lens being stained with fecal matter or opaque fluid. The presence of a colon fold and fecal matter or methods for fluid detection are used to indicate abnormal protrusions that may be haustral folds and creases or fecal residue. Finally, polyp detection is used to identify polyp protrusions in the colon lumen [3 (link)]. All of these functions are provided by the proposed CNN deep learning model. Figure 3 and Table 5 present the proposed CNN deep learning model architecture for the detection of blurred images, fecal matter, opaque water, and colon folds. Figure 4 and Table 6 present the polyp detection architecture for feature extraction and the bounding box transformation layer for the result output. Notably, polyp detection was performed on images from the six videos to verify the effectiveness of the system in identifying false alerts after low-quality images were excluded.
The size of each input image was measured in terms of the width (W) × Height (H) × filter number (N). All of the images were adjusted to fit the specification of the CNN deep learning model. We employed convolution, batch normalization, a rectified linear unit, and maximum pooling operations to conduct feature extraction [3 (link)]. Table 5 and Table 6 show the filters, size/stride, and output image size of each operation. A classification conversion layer was used to distinguish blurred images from non-polyp foreign body images. In addition, fully connected, softmax, and classification output layers were used for classification.