Mayur Mankar

Accepted at NAMSCON 2023

Digital health methodologies like A.I. with deep neural networks have become well-utilized for neuroimaging analysis tasks like segmenting brain lesions, atrophy or tumor detection, diagnosis and grading. However, recent diffident experiences demonstrate that carefully-engineered adversarial attacks can jeopardize medical informatics procedures and compromise imaging-based deep learning systems with small malicious imperceptible perturbations. This raises security concerns about the deployment of these computational neuroimaging systems in clinical settings. Our study is the first of its kind to investigate the robustness of deep learning-based MRI diagnostic systems using adversarial intrusions and formulates an iterative adversarial training approach to remedy these attacks.

Accepted at NAMSCON 2023

Evaluation of the neurocognitive salience level is an important critical need, whether in clinical monitoring or stressful occupational settings. Since high-resolution low-cost telemetric dermal-patch leads are available for physiological investigations, a suitable characteristic biopotential signal could be from an electrocoticographic/ electroencephalographic framework. It is well-known that in neurophysiological perspective, a cardinal parameter for stimulus-response characterization is the transinformation interaction flow (in hartleys/bits) which characterizes the information flow congruence between two states of the neurophysiological system. We have earlier transinformation measure to accurately predict the stimulus-response function of the neural system in animal preparations, and formulate a general neurocomputational basis of neural information transmission[1]. Here we investigate whether the transinformation parameter can also be utilized to delineate the neurocognitive salience level in different clinical situations.

Accepted at NAMSCON 2023

Digital health methodologies like A.I. with deep neural networks have become well-utilized for neuroimaging analysis tasks like segmenting brain lesions, atrophy or tumor detection, diagnosis and grading. However, recent diffident experiences demonstrate that carefully-engineered adversarial attacks can jeopardize medical informatics procedures and compromise imaging-based deep learning systems with small malicious imperceptible perturbations. This raises security concerns about the deployment of these computational neuroimaging systems in clinical settings. Our study is the first of its kind to investigate the robustness of deep learning-based MRI diagnostic systems using adversarial intrusions and formulates an iterative adversarial training approach to remedy these attacks.