senthil ramamurthy | Emory University (original) (raw)
Papers by senthil ramamurthy
Medical Image Analysis
Surgical tool detection is attracting increasing attention from the medical image analysis commun... more Surgical tool detection is attracting increasing attention from the medical image analysis community. The goal generally is not to precisely locate tools in images, but rather to indicate which tools are being used by the surgeon at each instant. The main motivation for annotating tool usage is to design efficient solutions for surgical workflow analysis, with potential applications in report generation, surgical training and even real-time decision support. Most existing tool annotation algorithms focus on laparoscopic surgeries. However, with 19 million interventions per year, the most common surgical procedure in the world is cataract surgery. The CATARACTS challenge was organized in 2017 to evaluate tool annotation algorithms in the specific context of cataract surgery. It relies on more than nine hours of videos, from 50 cataract surgeries, in which the presence of 21 surgical tools was manually annotated by two experts. With 14 participating teams, this challenge can be considered a success. As might be expected, the submitted solutions are based on deep learning. This paper thoroughly evaluates these solutions: in particular, the quality of their annotations are compared to that of human interpretations. Next, lessons learnt from the differential analysis of these solutions are discussed. We expect that they will guide the design of efficient surgery monitoring tools in the near future.
Journal of Cardiovascular Magnetic Resonance, 2016
Journal of Cardiovascular Magnetic Resonance, 2016
Physics in medicine and biology, Jan 7, 2016
A previously proposed x-ray scatter correction method for dedicated breast computed tomography wa... more A previously proposed x-ray scatter correction method for dedicated breast computed tomography was further developed and implemented so as to allow for initial patient testing. The method involves the acquisition of a complete second set of breast CT projections covering 360° with a perforated tungsten plate in the path of the x-ray beam. To make patient testing feasible, a wirelessly controlled electronic positioner for the tungsten plate was designed and added to a breast CT system. Other improvements to the algorithm were implemented, including automated exclusion of non-valid primary estimate points and the use of a different approximation method to estimate the full scatter signal. To evaluate the effectiveness of the algorithm, evaluation of the resulting image quality was performed with a breast phantom and with nine patient images. The improvements in the algorithm resulted in the avoidance of introduction of artifacts, especially at the object borders, which was an issue in...
Echocardiography, 2016
Tricuspid annular plane systolic excursion (TAPSE) has emerged as a reliable marker of right vent... more Tricuspid annular plane systolic excursion (TAPSE) has emerged as a reliable marker of right ventricular (RV) systolic function. Recently, TAPSE derived using 2D images (2D-TAPSE) was shown to correlate with M-mode TAPSE (MM-TAPSE). We have developed a novel technique for semiautomatic evaluation of TAPSE (SA-TAPSE). The purpose of this study was to determine the accuracy of this novel technique and validate it on normal hearts and pulmonary hypertension (PH). A total of 110 patients (56 with normal heart and 54 with PH) were retrospectively identified for analysis. The semiautomatic algorithm tracked the lateral tricuspid valve hinge point (TVHP) and the apex in the apical 4-chamber view. SA-TAPSE was calculated as displacement of the TVHP in end-diastole (ED) and end-systole (ES). The same points were manually identified to derive 2D-TAPSE. The system was able to accurately identify ED and ES in 304/330 heartbeats within three cardiac frames. The automatically identified TVHP points were within 1.2 ± 0.7 mm from the manually identified points. Intra-class correlation between SA-TAPSE and 2D-TAPSE was 0.96 (95% CI 0.93-0.98) for normal hearts and 0.92 (95% CI 0.87-0.96) for those with PH. Bland-Altman analysis showed a strong agreement between SA-TAPSE and 2D-TAPSE for normal hearts and those with PH. Using the novel custom-made software, SA-TAPSE could be measured in majority of our patients and was accurate when applied to normal hearts and those with PH. Future work will focus on fully automating the system for a rapid retrospective analysis of TAPSE.
Critical Care Medicine, 2015
Journal of Cardiovascular …, 2010
We introduce the concept, benefits, and general architecture for acquiring, storing, and displayi... more We introduce the concept, benefits, and general architecture for acquiring, storing, and displaying digital photographs along with medical imaging examinations. We also discuss a specific implementation built around an Androidbased system for simultaneously acquiring digital photographs along with portable radiographs. By an innovative application of radiofrequency identification technology to radiographic cassettes, the system is able to maintain a tight relationship between these photographs and the radiographs within the picture archiving and communications system (PACS) environment. We provide a cost analysis demonstrating the economic feasibility of this technology. Since our architecture naturally integrates with patient identification methods, we also address patient privacy issues.
A well-known reconstruction technique developed by Gerchberg, based on “error energy” reduction [... more A well-known reconstruction technique developed by Gerchberg, based on “error energy” reduction [1], is extended in this study to sparsely sampled dynamic cardiac magnetic resonance imaging (CMRI). A-priori knowledge of static and dynamic regions in the FOV is used to sample a subset of phase-encoding views on a regular Cartesian grid, allowing a reduction in overall imaging time. Similar to the direct-inversion Noquist method [2], the iterative reconstruction does not use either data-substitution or temporal interpolation. ...
2008 15th IEEE International Conference on Image Processing, 2008
A comparison study is presented of two methods for MRI reconstruction using super-resolution tech... more A comparison study is presented of two methods for MRI reconstruction using super-resolution techniques which combine multiple multi-slice stacks into a single high- resolution 3-D image volume. Sampling configurations are compared involving stacks with parallel orientations at different sub-pixel offset locations, and stacks with regularly distributed slice orientations. Results from experiments with simulated and real MRI data suggest that different
Journal of Digital Imaging, 2014
Our objective is to design, implement, and phantom-test a device to automatically obtain point-of... more Our objective is to design, implement, and phantom-test a device to automatically obtain point-of-care patient photographs along with portable radiographs. Such photographs could help with detection of wrong-patient errors. Our device consists of a camera controller (CC) and a camera that can be mounted on a portable conventional radiography (CR) machine. Radiation from the CR machine triggers an identification module (IM) embedded in the CR cassette. The IM then sends the cassette identifier-Plate_ID-to the CC along with a trigger to activate the camera. This trigger ensures simultaneous acquisition of radiograph and photograph, and the Plate_ID along with a time stamp ensures binding of the two images. We conducted phantom tests to determine if clinical portable radiography exposure settings (90 to 120 kVp and exposure time ranging from 0.63 to 8.0 ms) are sufficient to trigger the IM. Phantom experiments demonstrate acceptable sensor performance for clinical portable radiography exposures. Simultaneous acquisition of photographs is achieved by integrating a low-cost identification module containing a scintillator-detector into the radiographic cassette. Incident X-rays activate the scintillator-detector triggering photograph acquisition by a camera controller.
2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2014
Clinical decision support systems use image processing and machine learning methods to objectivel... more Clinical decision support systems use image processing and machine learning methods to objectively predict cancer in histopathological images. Integral to the development of machine learning classifiers is the ability to generalize from training data to unseen future data. A classification model's ability to accurately predict class label for new unseen data is measured by performance metrics, which also informs the classifier model selection process. Based on our research, commonly used metrics in literature (such as accuracy, ROC curve) do not accurately reflect the trained model's robustness. To the best of our knowledge, no research has been conducted to quantitatively compare performance metrics in the context of cancer prediction in histopathological images. In this paper, we evaluate various performance metrics and show that the Lift metric has the highest correlation between internal and external validation sets of a nested cross validation pipeline (R(2) = 0.57). Thus, we demonstrate that the Lift metric best generalizes classifier performance among the 23 metrics that were evaluated. Using the lift metric, we develop a classifier with a misclassification rate of 0.25 (4-class classifier) for data that the model was not trained on (external validation).
2008 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2008
This paper introduces a novel method for accelerated dynamic image acquisition for cardiac MRI. T... more This paper introduces a novel method for accelerated dynamic image acquisition for cardiac MRI. This method combines two different formalisms for reconstruction from sparse data by incorporation of prior information. Parallel imaging uses information about coil geometry in imaging systems with multiple receiver coils. Reduced field of view (rFOV) imaging exploits knowledge about static regions in a dynamic image scene. The novel method combines the SPACE-RIP implementation of parallel imaging with the Noquist rFOV imaging method, which both use a direct inversion model for image reconstruction. The theory is presented for integrated application of these methods, and results are presented of supporting experiments with simulated and real MRI data, retrospectively subsampled to generate sparse data sets. Successful application of the method confirms multiplicative combined acceleration.
Medical Physics, 2013
To analyze and optimize the signal-to-noise ratio (SNR) for the "Noquist" method for acceleration... more To analyze and optimize the signal-to-noise ratio (SNR) for the "Noquist" method for acceleration of cine magnetic resonance imaging in the presence of partially static field of view, designing practical methods for selection of optimal or near-optimal sample sets to allow reliable application of the method for variable image dimensions. Methods: To investigate the impact of the Noquist method and its experimental parameters on the SNR in the image reconstructed from reduced data, and to explore optimization of methods for highest SNR stability, three different optimization parameters have been selected: the condition of the forward matrix (R cond ) as it defines the propagation of noise into the reconstructed image, and the maximum ( maxD ) and the mean ( meanD ) linear noise amplification factor of the dynamic field-ofview (FOV) region. As SNR in a Noquist reconstruction is often not uniform across the FOV and since dynamic regions may contain the part of the image more clinically relevant, primarily these noise levels are targeted for optimization. Using these three optimization parameters, three experiments were conducted: characterization of Noquist SNR properties as a function of important image size parameters; for sufficiently small image dimensions, employment of exhaustive search using lexicographical algorithms to visit all possibilities under the cine imaging constraint that equal numbers of views are acquired at each time point of the sequence; and, departing from an hypothetically optimal pattern, generation and evaluation of SNR characteristics of a series of random variations to that optimal pattern. Results: The impact of favorable sparse data selection is illustrated, and SNR properties are characterized as a function of relevant acquisition parameters. Optimal data selection is investigated by exhaustive methods for small image sizes, and compared with algorithmic selection patterns. Observations from these experiments are confirmed by further studies on data selection for realistic image dimensions and an optimal selection algorithm is proposed. Sixty-four cases of small image sizes were analyzed through exhaustive search with a total of 527 984 141 matrix inversions called in the process, evaluating several SNR parameters for each case. An algorithm, named "Stairwell," that permits to design image dimensions with optimal SNR characteristics is presented, evaluated and compared with cases analyzed through exhaustive search. In 71.9% of the cases exhaustively studied, the Stairwell algorithm yielded optimal solutions. For no case did the deviation from optimum exceed 3.2% (R cond ), 1.0% ( meanD ), and 4.9% ( maxD ).
Magnetic Resonance in Medicine, 2011
This article introduces a novel method named ''Parallel Imaging and NOquist in Tandem'' (PINOT) f... more This article introduces a novel method named ''Parallel Imaging and NOquist in Tandem'' (PINOT) for accelerated image acquisition of cine cardiac magnetic resonance imaging. This method combines two prior information formalisms, the SPACE-RIP implementation of parallel imaging and the Noquist method for reduced-data image reconstruction with prior knowledge of static and dynamic regions in the field of view. The general theory is presented, and supported by results from experiments using time-resolved two-dimensional simulation data and retrospectively sub-sampled magnetic resonance imaging data with acceleration factors around 4. A signal-to-noise ratio analysis and a comparison study with TSENSE and k-t SENSE show that PINOT performs favorably in preserving edge detail, at a cost in signal-tonoise ratio and computational complexity. Magn Reson Med 65:1062-1075,
Medical Image Analysis
Surgical tool detection is attracting increasing attention from the medical image analysis commun... more Surgical tool detection is attracting increasing attention from the medical image analysis community. The goal generally is not to precisely locate tools in images, but rather to indicate which tools are being used by the surgeon at each instant. The main motivation for annotating tool usage is to design efficient solutions for surgical workflow analysis, with potential applications in report generation, surgical training and even real-time decision support. Most existing tool annotation algorithms focus on laparoscopic surgeries. However, with 19 million interventions per year, the most common surgical procedure in the world is cataract surgery. The CATARACTS challenge was organized in 2017 to evaluate tool annotation algorithms in the specific context of cataract surgery. It relies on more than nine hours of videos, from 50 cataract surgeries, in which the presence of 21 surgical tools was manually annotated by two experts. With 14 participating teams, this challenge can be considered a success. As might be expected, the submitted solutions are based on deep learning. This paper thoroughly evaluates these solutions: in particular, the quality of their annotations are compared to that of human interpretations. Next, lessons learnt from the differential analysis of these solutions are discussed. We expect that they will guide the design of efficient surgery monitoring tools in the near future.
Journal of Cardiovascular Magnetic Resonance, 2016
Journal of Cardiovascular Magnetic Resonance, 2016
Physics in medicine and biology, Jan 7, 2016
A previously proposed x-ray scatter correction method for dedicated breast computed tomography wa... more A previously proposed x-ray scatter correction method for dedicated breast computed tomography was further developed and implemented so as to allow for initial patient testing. The method involves the acquisition of a complete second set of breast CT projections covering 360° with a perforated tungsten plate in the path of the x-ray beam. To make patient testing feasible, a wirelessly controlled electronic positioner for the tungsten plate was designed and added to a breast CT system. Other improvements to the algorithm were implemented, including automated exclusion of non-valid primary estimate points and the use of a different approximation method to estimate the full scatter signal. To evaluate the effectiveness of the algorithm, evaluation of the resulting image quality was performed with a breast phantom and with nine patient images. The improvements in the algorithm resulted in the avoidance of introduction of artifacts, especially at the object borders, which was an issue in...
Echocardiography, 2016
Tricuspid annular plane systolic excursion (TAPSE) has emerged as a reliable marker of right vent... more Tricuspid annular plane systolic excursion (TAPSE) has emerged as a reliable marker of right ventricular (RV) systolic function. Recently, TAPSE derived using 2D images (2D-TAPSE) was shown to correlate with M-mode TAPSE (MM-TAPSE). We have developed a novel technique for semiautomatic evaluation of TAPSE (SA-TAPSE). The purpose of this study was to determine the accuracy of this novel technique and validate it on normal hearts and pulmonary hypertension (PH). A total of 110 patients (56 with normal heart and 54 with PH) were retrospectively identified for analysis. The semiautomatic algorithm tracked the lateral tricuspid valve hinge point (TVHP) and the apex in the apical 4-chamber view. SA-TAPSE was calculated as displacement of the TVHP in end-diastole (ED) and end-systole (ES). The same points were manually identified to derive 2D-TAPSE. The system was able to accurately identify ED and ES in 304/330 heartbeats within three cardiac frames. The automatically identified TVHP points were within 1.2 ± 0.7 mm from the manually identified points. Intra-class correlation between SA-TAPSE and 2D-TAPSE was 0.96 (95% CI 0.93-0.98) for normal hearts and 0.92 (95% CI 0.87-0.96) for those with PH. Bland-Altman analysis showed a strong agreement between SA-TAPSE and 2D-TAPSE for normal hearts and those with PH. Using the novel custom-made software, SA-TAPSE could be measured in majority of our patients and was accurate when applied to normal hearts and those with PH. Future work will focus on fully automating the system for a rapid retrospective analysis of TAPSE.
Critical Care Medicine, 2015
Journal of Cardiovascular …, 2010
We introduce the concept, benefits, and general architecture for acquiring, storing, and displayi... more We introduce the concept, benefits, and general architecture for acquiring, storing, and displaying digital photographs along with medical imaging examinations. We also discuss a specific implementation built around an Androidbased system for simultaneously acquiring digital photographs along with portable radiographs. By an innovative application of radiofrequency identification technology to radiographic cassettes, the system is able to maintain a tight relationship between these photographs and the radiographs within the picture archiving and communications system (PACS) environment. We provide a cost analysis demonstrating the economic feasibility of this technology. Since our architecture naturally integrates with patient identification methods, we also address patient privacy issues.
A well-known reconstruction technique developed by Gerchberg, based on “error energy” reduction [... more A well-known reconstruction technique developed by Gerchberg, based on “error energy” reduction [1], is extended in this study to sparsely sampled dynamic cardiac magnetic resonance imaging (CMRI). A-priori knowledge of static and dynamic regions in the FOV is used to sample a subset of phase-encoding views on a regular Cartesian grid, allowing a reduction in overall imaging time. Similar to the direct-inversion Noquist method [2], the iterative reconstruction does not use either data-substitution or temporal interpolation. ...
2008 15th IEEE International Conference on Image Processing, 2008
A comparison study is presented of two methods for MRI reconstruction using super-resolution tech... more A comparison study is presented of two methods for MRI reconstruction using super-resolution techniques which combine multiple multi-slice stacks into a single high- resolution 3-D image volume. Sampling configurations are compared involving stacks with parallel orientations at different sub-pixel offset locations, and stacks with regularly distributed slice orientations. Results from experiments with simulated and real MRI data suggest that different
Journal of Digital Imaging, 2014
Our objective is to design, implement, and phantom-test a device to automatically obtain point-of... more Our objective is to design, implement, and phantom-test a device to automatically obtain point-of-care patient photographs along with portable radiographs. Such photographs could help with detection of wrong-patient errors. Our device consists of a camera controller (CC) and a camera that can be mounted on a portable conventional radiography (CR) machine. Radiation from the CR machine triggers an identification module (IM) embedded in the CR cassette. The IM then sends the cassette identifier-Plate_ID-to the CC along with a trigger to activate the camera. This trigger ensures simultaneous acquisition of radiograph and photograph, and the Plate_ID along with a time stamp ensures binding of the two images. We conducted phantom tests to determine if clinical portable radiography exposure settings (90 to 120 kVp and exposure time ranging from 0.63 to 8.0 ms) are sufficient to trigger the IM. Phantom experiments demonstrate acceptable sensor performance for clinical portable radiography exposures. Simultaneous acquisition of photographs is achieved by integrating a low-cost identification module containing a scintillator-detector into the radiographic cassette. Incident X-rays activate the scintillator-detector triggering photograph acquisition by a camera controller.
2014 36th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, 2014
Clinical decision support systems use image processing and machine learning methods to objectivel... more Clinical decision support systems use image processing and machine learning methods to objectively predict cancer in histopathological images. Integral to the development of machine learning classifiers is the ability to generalize from training data to unseen future data. A classification model's ability to accurately predict class label for new unseen data is measured by performance metrics, which also informs the classifier model selection process. Based on our research, commonly used metrics in literature (such as accuracy, ROC curve) do not accurately reflect the trained model's robustness. To the best of our knowledge, no research has been conducted to quantitatively compare performance metrics in the context of cancer prediction in histopathological images. In this paper, we evaluate various performance metrics and show that the Lift metric has the highest correlation between internal and external validation sets of a nested cross validation pipeline (R(2) = 0.57). Thus, we demonstrate that the Lift metric best generalizes classifier performance among the 23 metrics that were evaluated. Using the lift metric, we develop a classifier with a misclassification rate of 0.25 (4-class classifier) for data that the model was not trained on (external validation).
2008 5th IEEE International Symposium on Biomedical Imaging: From Nano to Macro, 2008
This paper introduces a novel method for accelerated dynamic image acquisition for cardiac MRI. T... more This paper introduces a novel method for accelerated dynamic image acquisition for cardiac MRI. This method combines two different formalisms for reconstruction from sparse data by incorporation of prior information. Parallel imaging uses information about coil geometry in imaging systems with multiple receiver coils. Reduced field of view (rFOV) imaging exploits knowledge about static regions in a dynamic image scene. The novel method combines the SPACE-RIP implementation of parallel imaging with the Noquist rFOV imaging method, which both use a direct inversion model for image reconstruction. The theory is presented for integrated application of these methods, and results are presented of supporting experiments with simulated and real MRI data, retrospectively subsampled to generate sparse data sets. Successful application of the method confirms multiplicative combined acceleration.
Medical Physics, 2013
To analyze and optimize the signal-to-noise ratio (SNR) for the "Noquist" method for acceleration... more To analyze and optimize the signal-to-noise ratio (SNR) for the "Noquist" method for acceleration of cine magnetic resonance imaging in the presence of partially static field of view, designing practical methods for selection of optimal or near-optimal sample sets to allow reliable application of the method for variable image dimensions. Methods: To investigate the impact of the Noquist method and its experimental parameters on the SNR in the image reconstructed from reduced data, and to explore optimization of methods for highest SNR stability, three different optimization parameters have been selected: the condition of the forward matrix (R cond ) as it defines the propagation of noise into the reconstructed image, and the maximum ( maxD ) and the mean ( meanD ) linear noise amplification factor of the dynamic field-ofview (FOV) region. As SNR in a Noquist reconstruction is often not uniform across the FOV and since dynamic regions may contain the part of the image more clinically relevant, primarily these noise levels are targeted for optimization. Using these three optimization parameters, three experiments were conducted: characterization of Noquist SNR properties as a function of important image size parameters; for sufficiently small image dimensions, employment of exhaustive search using lexicographical algorithms to visit all possibilities under the cine imaging constraint that equal numbers of views are acquired at each time point of the sequence; and, departing from an hypothetically optimal pattern, generation and evaluation of SNR characteristics of a series of random variations to that optimal pattern. Results: The impact of favorable sparse data selection is illustrated, and SNR properties are characterized as a function of relevant acquisition parameters. Optimal data selection is investigated by exhaustive methods for small image sizes, and compared with algorithmic selection patterns. Observations from these experiments are confirmed by further studies on data selection for realistic image dimensions and an optimal selection algorithm is proposed. Sixty-four cases of small image sizes were analyzed through exhaustive search with a total of 527 984 141 matrix inversions called in the process, evaluating several SNR parameters for each case. An algorithm, named "Stairwell," that permits to design image dimensions with optimal SNR characteristics is presented, evaluated and compared with cases analyzed through exhaustive search. In 71.9% of the cases exhaustively studied, the Stairwell algorithm yielded optimal solutions. For no case did the deviation from optimum exceed 3.2% (R cond ), 1.0% ( meanD ), and 4.9% ( maxD ).
Magnetic Resonance in Medicine, 2011
This article introduces a novel method named ''Parallel Imaging and NOquist in Tandem'' (PINOT) f... more This article introduces a novel method named ''Parallel Imaging and NOquist in Tandem'' (PINOT) for accelerated image acquisition of cine cardiac magnetic resonance imaging. This method combines two prior information formalisms, the SPACE-RIP implementation of parallel imaging and the Noquist method for reduced-data image reconstruction with prior knowledge of static and dynamic regions in the field of view. The general theory is presented, and supported by results from experiments using time-resolved two-dimensional simulation data and retrospectively sub-sampled magnetic resonance imaging data with acceleration factors around 4. A signal-to-noise ratio analysis and a comparison study with TSENSE and k-t SENSE show that PINOT performs favorably in preserving edge detail, at a cost in signal-tonoise ratio and computational complexity. Magn Reson Med 65:1062-1075,