Machine Learning Based Autism Detection using Brain MRI (PhD Research)



I am making an effort to better understand the autistic brain and identify the neuroanatomical basis of autism. I apply computer vision and machine learning techniques on brain images to extract valuable insights. Currently, my research focuses on data-driven discovery of brain biomarkers for early detection of autism.

Related Publications:

  1. G. J. Katuwal,  Machine Learning Based Autism Detection Using Brain Imaging, PhD thesis, Rochester Institute of Technology, May 2017 [link]
  2. G. J. Katuwal, N. D. Cahill, Chase. C. Doughetry, S. Baum, G. J. Moore, Eli Evans, A. M. Michael, Inter-method discrepancies in brain volume estimation may drive inconsistent findings in autismFrontiers in Neuroscience, 10:439, September 2016 [link]
  3. C. C. Dougherty, D. W. Evans, G. J. Katuwal, A. M. Michael, Asymmetry of fusiform structure in autism spectrum disorder: Trajectory and association with symptom severity, Molecular Autism, 7(1):1, May 2016 [link]
  4. G. J. Katuwal, S. Baum, N. D. Cahill, A. M. Michael, Divide and Conquer: Sub-Grouping of ASD Improves ASD Detection, PLoS ONE, 11(4):e0153331, April 2016 [link]
  5. G. J. Katuwal, N. D. Cahill, S. Baum, A. M. Michael, The Predictive Power of Structural MRI in Autism Diagnosis, IEEE EMBS, Italy, August 2015 [link]

Patient Level Model Interpretability for Precision Medicine

embc2016_precision_medicine_poster_figureInterpretability of machine learning models is critical for precision medicine efforts. However, highly predictive models are generally complex and are difficult to interpret. Here using Local Interpretable Model-agnostic Explanations (LIME) algorithm, we show that complex models such as random forest can be made interpretable. In MIMIC-II dataset, we achieved a high mortality prediction rate of ICU patients and were also able to identify the contribution of the important features for prediction at the patient level. We achieved this by combining the predictive power from RF with the patient-level model interpretability from LIME, where we linearly approximated the RF model in the patient vicinity. The interpretation from the LIME model for our test patient is consistent with current medical knowledge.

Related Publications:

  1. G. J. Katuwal, R. Chen, Machine Learning Model Interpretability for Precision Medicine, IEEE EMBS, Orlando, August 2016 [link]

Automatic Fundus Image Field Detection and Quality Assessment

Funded by CIS Research micro-grant $6700

We designed a framework to provide the immediate feedback on the acquired fundus images. Our system is able to detect the field and side of the fundus images with 99% accuracy. The system was also able to predict the quality of individual fundus images as well as the overall quality from a set of fundus images.

Related Publications:

  1. G. J. Katuwal, J. P. Kerekes, R. Ramchandran, C. Sisson and N. Rao, Automatic Fundus Image Field Detection and Quality Assessment, IEEE WNYIPW, Rochester, NY, November 2013
  2. G.J. Katuwal, J.P. Kerekes, R. Ramchandran, C. Sisson, Automatic Fundus Image Field Detection and Quality Assessment, U.S. Patent Application 14/511936, Issued April 2015