Computer-aided diagnosis of colorectal polyps using linked color imaging colonoscopy to predict histology

Abstract

We developed a computer-aided diagnosis (CAD) system based on linked color imaging (LCI) images to predict the histological results of polyps by analyzing the colors of the lesions. A total of 139 images of adenomatous polyps and 69 images of non-adenomatous polyps obtained from our hospital were collected and used to train the CAD system. A test set of LCI images, including both adenomatous and non-adenomatous polyps, was prospectively collected from patients who underwent colonoscopies between Oct and Dec 2017; this test set was used to assess the diagnostic abilities of the CAD system compared to those of human endoscopists (two experts and two novices). The accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of this novel CAD system for the training set were 87.0%, 87.1%, 87.0%, 93.1%, and 76.9%, respectively. The test set included 115 adenomatous polyps and 66 non-adenomatous polyps that were prospectively collected. The CAD system identified adenomatous or non-adenomatous polyps in the test set with an accuracy of 78.4%, a sensitivity of 83.3%, a specificity of 70.1%, a PPV of 82.6%, and an NPV of 71.2%. The accuracy of the CAD system was comparable to that of the expert endoscopists (78.4% vs 79.6%; p = 0.517). In addition, the diagnostic accuracy of the novices was significantly lower to the performance of the experts (70.7% vs 79.6%; p = 0.018). A novel CAD system based on LCI could be a rapid and powerful decision-making tool for endoscopists.

Conflict of interest statement

This work was supported by the Beijing Nova Program (No. Z171100001117090).

Figures

Figure 1

Flowchart of the training process. The original LCI picture is preprocessed to a 64 × 64 image where each pixel is represented as a 3-D vector (with R value, G value and B value). Transform the RGB color space to HLS. Now each pixel is represented as a different 3-D vector with H value, L value and S value. Concatenate 2 vectors to a 6-D vector. Do concatenation at every pixel of the image then we have a 64 × 64 × 6 pixels block. Feed these blocks into two initialized models separately to train for two independent GMMs.

Figure 2

Flowchart of the threshold calculation. After training we obtain two GMMs. Transform all training images to pixels blocks and input the blocks to GMMs. We get two scores for each block consequently. Record the difference between the two scores of all blocks and plot the training ROC curve. Choose the point where the sensitivity is approximately equal to the specificity as the Threshold.

Figure 3

Flowchart of the test process. When a new image comes we transform it to pixels block and input it to two GMMs and obtain two scores. Get the difference between these two scores and compare it with the Threshold. If the difference is larger than Threshold, the image if classified as adenomatous image. Otherwise it is inflammatory image.

Figure 4

Receiver operator characteristic curve for the CAD differentiation of adenomatous versus hyperplastic polyps in the training set. AUC, area under the curve.