Systems and methods for digitizing electrocardiograms
The disclosure relates to systems and methods of converting a representation of a physiological signal (e.g., a non-digitized version such as a printed curve) into a digitized representation of the physiological signal of a subject. For example, a printed electrocardiogram (ECG) may be digitized using the systems are methods provided herein. The method may include receiving a digitized image of a printed curve representing the physiological signal of the subject, and detecting at least one region of interest having a portion of the physiological signal. For each of the regions of interest, the method may include extracting coordinates representing the physiological signal and registering the extracted coordinates.
1. A method for digital conversion of a printed representation of a physiological signal of a subject, the method comprising:
receiving a digitized image of the printed representation of the physiological signal of the subject;
detecting, via a first neural network, a layout region of the digitized image and at least one sub-region dividing the layout region;
detecting, via a second neural network, at least one region of interest inside one of the identified sub-regions comprising a portion of the physiological signal; and
extracting, via a third neural network, coordinates representing the physiological signal for each of the at least one region of interest.
2. The method of claim 1 , wherein the physiological signal includes an electrocardiogram.
3. The method of claim 1 , further comprising registering the coordinates representing the physiological signal on a computer readable storage medium.
4. The method of claim 1 , wherein the first neural network receives as an input the digitized image and generates as an output one or more of (i) at least one characteristic dimension, (ii) a position of the layout region, and (iii) the number of sub-regions.
5. The method of claim 1 , wherein the second neural network receives as an input an image of one of the at least one sub-region and generates as an output one or more of (i) a dimension and a position of the at least one region of interest, and (ii) a probability of the presence of a portion of the physiological signal in the at least one region of interest.
6. The method of claim 1 , wherein the third neural network receives as an input one or more of (i) a dimension and a position of the at least one region of interest, and (ii) a probability of the presence of a portion of the physiological signal in the at least one region of interest and generates as an output the coordinates representing the physiological signal.
7. The method of claim 6 , wherein the third neural network generates as an output a probability map used to obtain the coordinates representing the physiological signal.
8. The method of claim 1 , wherein the first neural network comprises at least two hidden layers, the second neural network comprises at least two hidden layers, and the third neural network comprises at least two hidden layers.
9. The method of claim 1 , wherein the second neural network includes a convolutional neural network and comprises a pooling layer.
10. The method of claim 1 , wherein the third neural network is a convolutional neural network comprising at least one convolutional layer followed by at least one pooling layer and at least one transpose convolutional layer.
11. A method for digital conversion of a printed representation of a physiological signal of a subject, the method comprising:
receiving a digitized image of a printed representation of a physiological signal;
detecting a layout region;
dividing the layout region into at least one sub-region;
generating at least one characteristic dimension of the layout region, a position of the layout region, and a number sub-regions;
for each of the at least one sub-regions, segmenting, via a segmentation neural network, at least one region of interest comprising a portion of the physiological signal, wherein an input of the segmentation neural network is an image of the at least one sub-region and an output comprises a dimensions and a position of the at least one region of interest, and a probability of the presence of a portion of the physiological signal in the at least one region of interest; and
for each region of interest, extracting, via an extraction neural network, coordinates representing the physiological signal.
12. The method of claim 11 , wherein the detection of the layout region and of the at least one sub-region is performed using a division neural network.
13. The method of claim 12 , wherein the division neural network comprises at least two hidden layers.
14. The method of claim 12 , wherein the division neural network comprises a convolutional neural network comprising at least two parallel dense layers, wherein a first parallel dense layer corresponds to the output of the characteristic dimension of the layout region and a second parallel dense layer corresponds to the number of sub-regions.
15. The method of claim 11 , wherein the segmentation neural network comprises at least two hidden layers and the extraction neural network comprises at least two hidden layers.
16. The method of claim 11 , wherein the output of the segmentation neural network is an input of the extraction neural network.
17. The method of claim 11 , further comprising, for each region of interest, generating, via an extraction neural network, a probability map from which the coordinates representing the physiological signal is extracted.
18. The method of claim 11 , wherein the segmentation neural network includes a convolutional neural network and comprises a pooling layer.
19. The method of claim 11 , wherein the extraction neural network is a convolutional neural network comprising at least one convolutional layer followed by at least one pooling layer and at least one transpose convolutional layer.
20. A system for conversion of a printed curve representing a physiological signal of a subject into a digitized curve, the system comprising instructions stored on at least one processor, the instructions configured to, when executed, cause the at least one processor to:
receive a digitized image of a printed curve representing the physiological signal of the subject;
detect, via a first neural network, a layout region of the digitized image and at least one sub-region dividing the layout region;
detect, via a second neural network, at least one region of interest comprising a portion of the physiological signal inside one of the at least one sub-region; and
extract, via a third neural network, coordinates representing the physiological signal for each of the at least one region of interest.