User login
The system, called CXR-Net, was trained to differentiate SARS-CoV-2 chest x-rays (CXRs) from CXRs that are either normal or non–COVID-19 lung pathologies, explained Abdulah Haikal, an MD candidate at Wayne State University, Detroit.
Mr. Haikal described CXR-Net at the AACR Virtual Meeting: COVID-19 and Cancer (Abstract S11-04).
CXR-Net is a two-module pipeline, Mr. Haikal explained. Module I is based on Res-CR-Net, a type of neural network originally designed for the semantic segmentation of microscopy images, with the ability to retain the original resolution of the input images in the feature maps of all layers and in the final output.
Module II is a hybrid convolutional neural network in which the first convolutional layer with learned coefficients is replaced by a layer with fixed coefficients provided by the Wavelet Scattering Transform. Module II inputs patients’ CXRs and corresponding lung masks quantified by Module I, and generates as outputs a class assignment (COVID-19 or non–COVID-19) and high-resolution heat maps that detect the severe acute respiratory syndrome–-associated lung regions.
“The system is trained to differentiate COVID and non-COVID pathologies and produces a highly discriminative heat map to point to lung regions where COVID is suspected,” Mr. Haikal said. “The Wavelet Scattering Transform allows for fast determination of COVID versus non-COVID CXRs.”
Preliminary results and implications
CXR-Net was piloted on a small dataset of CXRs from non–COVID-19 and polymerase chain reaction–confirmed COVID-19 patients acquired at a single center in Detroit.
Upon fivefold cross validation of the training set with 2,265 images, 90% accuracy was observed when the training set was tested against the validation set. However, once 1,532 new images were introduced, a 76% accuracy rate was observed.
The F1 scores were 0.81 and 0.70 for the training and test sets, respectively.
“I’m really excited about this new approach, and I think AI will allow us to do more with less, which is exciting,” said Ross L. Levine, MD, of Memorial Sloan Kettering Cancer Center in New York, who led a discussion session with Mr. Haikal about CXR-Net.
One question raised during the discussion was whether the technology will help health care providers be more thoughtful about when and how they image COVID-19 patients.
“The more data you feed into the system, the stronger and more accurate it becomes,” Mr. Haikal said. “However, until we have data sharing from multiple centers, we won’t see improved accuracy results.”
Another question was whether this technology could be integrated with more clinical parameters.
“Some individuals are afraid that AI will replace the job of a professional, but it will only make it better for us,” Mr. Haikal said. “We don’t rely on current imaging techniques to make a definitive diagnosis, but rather have a specificity and sensitivity to establish a diagnosis, and AI can be used in the same way as a diagnostic tool.”
Mr. Haikal and Dr. Levine disclosed no conflicts of interest. No funding sources were reported in the presentation.
The system, called CXR-Net, was trained to differentiate SARS-CoV-2 chest x-rays (CXRs) from CXRs that are either normal or non–COVID-19 lung pathologies, explained Abdulah Haikal, an MD candidate at Wayne State University, Detroit.
Mr. Haikal described CXR-Net at the AACR Virtual Meeting: COVID-19 and Cancer (Abstract S11-04).
CXR-Net is a two-module pipeline, Mr. Haikal explained. Module I is based on Res-CR-Net, a type of neural network originally designed for the semantic segmentation of microscopy images, with the ability to retain the original resolution of the input images in the feature maps of all layers and in the final output.
Module II is a hybrid convolutional neural network in which the first convolutional layer with learned coefficients is replaced by a layer with fixed coefficients provided by the Wavelet Scattering Transform. Module II inputs patients’ CXRs and corresponding lung masks quantified by Module I, and generates as outputs a class assignment (COVID-19 or non–COVID-19) and high-resolution heat maps that detect the severe acute respiratory syndrome–-associated lung regions.
“The system is trained to differentiate COVID and non-COVID pathologies and produces a highly discriminative heat map to point to lung regions where COVID is suspected,” Mr. Haikal said. “The Wavelet Scattering Transform allows for fast determination of COVID versus non-COVID CXRs.”
Preliminary results and implications
CXR-Net was piloted on a small dataset of CXRs from non–COVID-19 and polymerase chain reaction–confirmed COVID-19 patients acquired at a single center in Detroit.
Upon fivefold cross validation of the training set with 2,265 images, 90% accuracy was observed when the training set was tested against the validation set. However, once 1,532 new images were introduced, a 76% accuracy rate was observed.
The F1 scores were 0.81 and 0.70 for the training and test sets, respectively.
“I’m really excited about this new approach, and I think AI will allow us to do more with less, which is exciting,” said Ross L. Levine, MD, of Memorial Sloan Kettering Cancer Center in New York, who led a discussion session with Mr. Haikal about CXR-Net.
One question raised during the discussion was whether the technology will help health care providers be more thoughtful about when and how they image COVID-19 patients.
“The more data you feed into the system, the stronger and more accurate it becomes,” Mr. Haikal said. “However, until we have data sharing from multiple centers, we won’t see improved accuracy results.”
Another question was whether this technology could be integrated with more clinical parameters.
“Some individuals are afraid that AI will replace the job of a professional, but it will only make it better for us,” Mr. Haikal said. “We don’t rely on current imaging techniques to make a definitive diagnosis, but rather have a specificity and sensitivity to establish a diagnosis, and AI can be used in the same way as a diagnostic tool.”
Mr. Haikal and Dr. Levine disclosed no conflicts of interest. No funding sources were reported in the presentation.
The system, called CXR-Net, was trained to differentiate SARS-CoV-2 chest x-rays (CXRs) from CXRs that are either normal or non–COVID-19 lung pathologies, explained Abdulah Haikal, an MD candidate at Wayne State University, Detroit.
Mr. Haikal described CXR-Net at the AACR Virtual Meeting: COVID-19 and Cancer (Abstract S11-04).
CXR-Net is a two-module pipeline, Mr. Haikal explained. Module I is based on Res-CR-Net, a type of neural network originally designed for the semantic segmentation of microscopy images, with the ability to retain the original resolution of the input images in the feature maps of all layers and in the final output.
Module II is a hybrid convolutional neural network in which the first convolutional layer with learned coefficients is replaced by a layer with fixed coefficients provided by the Wavelet Scattering Transform. Module II inputs patients’ CXRs and corresponding lung masks quantified by Module I, and generates as outputs a class assignment (COVID-19 or non–COVID-19) and high-resolution heat maps that detect the severe acute respiratory syndrome–-associated lung regions.
“The system is trained to differentiate COVID and non-COVID pathologies and produces a highly discriminative heat map to point to lung regions where COVID is suspected,” Mr. Haikal said. “The Wavelet Scattering Transform allows for fast determination of COVID versus non-COVID CXRs.”
Preliminary results and implications
CXR-Net was piloted on a small dataset of CXRs from non–COVID-19 and polymerase chain reaction–confirmed COVID-19 patients acquired at a single center in Detroit.
Upon fivefold cross validation of the training set with 2,265 images, 90% accuracy was observed when the training set was tested against the validation set. However, once 1,532 new images were introduced, a 76% accuracy rate was observed.
The F1 scores were 0.81 and 0.70 for the training and test sets, respectively.
“I’m really excited about this new approach, and I think AI will allow us to do more with less, which is exciting,” said Ross L. Levine, MD, of Memorial Sloan Kettering Cancer Center in New York, who led a discussion session with Mr. Haikal about CXR-Net.
One question raised during the discussion was whether the technology will help health care providers be more thoughtful about when and how they image COVID-19 patients.
“The more data you feed into the system, the stronger and more accurate it becomes,” Mr. Haikal said. “However, until we have data sharing from multiple centers, we won’t see improved accuracy results.”
Another question was whether this technology could be integrated with more clinical parameters.
“Some individuals are afraid that AI will replace the job of a professional, but it will only make it better for us,” Mr. Haikal said. “We don’t rely on current imaging techniques to make a definitive diagnosis, but rather have a specificity and sensitivity to establish a diagnosis, and AI can be used in the same way as a diagnostic tool.”
Mr. Haikal and Dr. Levine disclosed no conflicts of interest. No funding sources were reported in the presentation.
FROM AACR: COVID-19 AND CANCER 2021