Aims Cecum intubation is the key criterion for the quality of colonoscopy. We tested the ability of computer-assisted image analysis using convolutional neural networks (CNNs) to confirm the achievement of the cecum.
Tab.1
Results of evaluating algorithm.
|
Validation set
|
Test set
|
|
ROC AUC
|
0.98
|
0.90
|
|
F1-score
|
0.873
|
0.886
|
Methods We designed and trained deep CNNs to detect the orifice of the appendix using a diverse and representative dataset of 1696 hand-labeled images from screening colonoscopies collected from more than 200 patients. This database was split in two datasets: training and validation with the ratio of 80% and 20%. Thus, 1356 images were used for training, which were additionally augmented using the geometric transformations. To evaluate the results during training, the validation set was used. In addition, the test set of 104 images was collected manually for the final evaluation of the results.We used a ResNet-50 network, which was pre-trained on the ImageNet dataset, and the output layer was replaced by three fully connected layers with dropout.
Results We used AUC and F1-score (the harmonic mean of the precision and recall) as an evaluation metrics. The threshold for computing F1-score was chosen optimal by the ROC-curve (receiver operating characteristic curve). The results are presented in the table.
As can be seen from the table the values of evaluation metrics are close for the validation and test sets, which indicates the strong generalization ability of the algorithm.
Conclusions The CNN identified the appendix orifice with AUC of 0.98 and F1-score of 0.873. The CNN system detected the appendix orifice in real-time constraints using an ordinary desktop machine with a contemporary graphics processing unit. This system could carry out an automated control for the quality of colonoscopy
