ECE PhD Proposal Review: Md Navid Akbar

PhD Proposal Review: Variational and Siamese Models in Functional and Structural Medical Image Analysis

Md Navid Akbar

Location: Zoom Link

Abstract: Machine learning (ML) models have recently shown great promise in medical image analysis. Instead of a one-size-fits-all, a customized model is generally needed to map a target outcome from an imaging modality. To this end, this proposal presents three such supervised models developed for three different imaging modalities.
In the first, a deep convolutional neural network (CNN) maps 3D cortical motor representation, obtained by transcranial magnetic stimulation (TMS), to the corresponding motor evoked potentials captured by surface electromyography (EMG). This modeling is bi-directional: with trivial changes, it can operate in both the forward and inverse directions. TMS as a functional imaging technique is still in its infancy, but its potential application in presurgical planning necessitates a reliable data-driven model. Our variational autoencoder inspired CNN is a pioneering step in that direction: with a normalized root mean square error up to below 14%, and an R-squared similarity up to above 87%, for cortical representation reconstruction in the inverse path. As the next steps, we plan to investigate other training strategies and collect additional data to assess robustness.
In the second, a Siamese CNN (with a pretrained DenseNet121 backbone) is developed to predict the continuous spectrum of pulmonary edema severity, from frontal chest X-rays. While existing deep learning frameworks have been promising in detecting the presence or absence of such edema, or even its discrete grades of severity, prediction of the continuous-valued severity remains a challenge. Using lower resolution images and only 1/51-th the size of training data compared to the state-of-the-art, our work beats it by achieving a mean area under the receiver operating characteristic curve (AUC) score of 91% (improvement by 4%), when tested on the open-source MIMIC-CXR database.
Finally, a complete preprocessing and ML classification pipeline is developed for identifying which traumatic brain injury (TBI) patients will go on to develop late seizures, from diffusion-weighted MRI (dMRI). Physical deformations following moderate-severe TBI present problems for standard processing of dMRI, complicating the extraction of neuroimaging features. Following the novel application of a normalization technique to dMRI, in conjunction with univariate feature selection and a linear discriminant analysis classifier, our model improves the performance over the standard pipeline by 8% in mean accuracy and 7% in mean AUC. In future work, we would like to explore classification using a fusion of dMRI with electroencephalogram (EEG) and functional MRI (fMRI) modalities.