Automatic detection of colon lesions and blood in colon capsule endoscopy
The present invention relates to a computer-implemented method, capable of automatically detecting clinically relevant colonic pleomorphic lesions and blood or hematic traces in colon capsule endoscopy images, by classifying pixels as colonic pleomorphic lesions or blood or hematic traces, using a convolutional image feature extraction step followed by a classification and indexing step of such findings into a set of one or more classes.
1 . A computer-implemented method capable of automatically detecting clinically relevant pleomorphic colonic lesions and blood or hematic residues in medical images, comprising:
detecting colonic lesions and blood or hematic residues in a medical images by classifying pixels as pleomorphic colonic lesions or blood or hematic residues using a convolutional image feature extraction step followed by a classification step and indexing such findings in a set of one of more classes wherein such method comprises:
the pre-training of a sample of said images in a plurality of combinations of said convolutional image feature extraction step followed by a classification component architecture using pre trained images as training data;
training of said images with the architecture combination that performs the best;
prediction of pleomorphic colonic lesions and blood traces using said optimized architecture combination;
including an output collector with means of communication to third-party validation capable of interpreting the accuracy of the previous steps of the method and correcting a wrong prediction;
storing the corrected prediction into the storage component.
2 . The method of claim 1 , wherein the medical images are CCE images.
3 . The method of claim 1 , wherein the pre-training runs iteratively in an optimization loop.
4 . The method of claim 1 , wherein the classification network architecture comprises at least two blocks, each having, a Dense layer followed by a Dropout layer.
5 . The method of claim 4 , wherein the last block of the classification component includes a BatchNormalization layer followed by a Dense layer with the depth size equal to the number of lesions type one wants to classify.
6 . The method of claim 1 , wherein the set of pre-trained neural networks is the best performing of the following: VGG16, InceptionV3, Xception, EfficientNetB5, EfficientNetB7, Resnet50 or Resnet125.
7 . The method of claim 6 , wherein the best-performing architecture is chosen based on the overall accuracy and f1-metrics.
8 . The method of claim 6 , wherein the training of the best-performing architecture comprises two to four dense layers in sequence, starting with 4096 and decreasing in half up to 512.
9 . The method of claim 8 , wherein between the final two layers of the best performing network there is a dropout layer of 0.1 drop rate.
10 . The method of claim 6 , wherein between the final two layers of the best performing network there is a dropout layer of 0.1 drop rate.
11 . The method of claim 1 , wherein the training of images includes the division in one or more patient exclusive folds for optimized training.
12 . The method of claim 11 , wherein the training of the images includes a ratio of training-to-validation of 10%-90%.
13 . The method of claim 12 , wherein the lesions are classified as colonic lesions or luminal blood traces.
14 . The method of claim 1 , wherein the third-party validation is done by user-input.
15 . The method of claim 14 , wherein the training dataset includes images in the storage component that were predicted by sequentially performing the steps of such method.
16 . A portable endoscopic device comprising instructions which, when executed by a processor, cause the computer to carry out the steps of the method of claim 1 .
17 . The method of claim 1 , wherein the last block of the classification component includes a BatchNormalization layer followed by a Dense layer with the depth size equal to the number of lesions type one wants to classify.
18 . The method of claim 1 , wherein the best-performing architecture is chosen based on the overall accuracy and f1-metrics.
19 . The method of claim 1 , wherein the training of the best-performing architecture comprises two to four dense layers in sequence, starting with 4096 and decreasing in half up to 512.
20 . The method of claim 1 , wherein between the final two layers of the best performing network there is a dropout layer of 0.1 drop rate.
21 . The method of claim 1 , wherein the training of the images includes a ratio of training-to-validation of 10%-90%.
22 . The method of claim 1 , wherein the lesions are classified as colonic lesions or luminal blood traces.
23 . The method of claim 1 , wherein the training dataset includes images in the storage component that were predicted by sequentially performing the steps of such method.