Page 48 - Biennial Report 2018-20 Jun 2021

A collaborative project of Souvik Maiti and Mary Ekka focused on the study of the lncRNA
MALAT-1 interaction with the protein nucleolin using biophysical approaches. MALAT-1 has
three putative G-quadruplex sequences (PQRS) present at its 3' end and these motifs were
assessed for their ability to form G-quadruplexes. Pull-down studies with the PQRS showed that
nucleolin was a potential interacting partner with MALAT-1. Cloned, expressed and purified
recombinant human (delta)nucleolin (minus 283 amino acids from the N-terminal) was used to
perform preliminary real time binding studies with different PQRS using surface plasmon
resonance. Preliminary data suggests that all the three G-quadruplexes bind to nucleolin, albeit
with differential affinity.

RAPID-CT: RADIOLOGICAL AI SYSTEM FOR PARALLEL INFORMATIC DETECTION
OF CLINICAL TRIAGE EMERGENCIES

Radiology and machine learning are bound to be working together in the near future given the
increasing rate of computational power and radiological patient workload. With the increasing
patient count and the effort required to master radiology, the increase in the number of
radiologists falls short of the increase in the number of patients. To aid the radiologist in making
quicker decisions and better reports, the RAPID-CT project proposes a system that enhances and
augments the radiologist workflow by using artificial intelligence to prioritize patients based on
risk score and annotate the scans to improve turnaround time for needy patients.
More than 130 samples were collected under the guidance of an expert radiologist (0.0625mm
to 0.5mm slice width, varying dimensions with a mode of 512x512 pixels) that were anonymized
(using in-house code) and transferred to the lab server via a secure FTP channel. The freely
available dataset from Qure.ai (497 patients, 0.0625 to 0.5mm slice width, 512x512 pixel
dimensions) was also collected.
Exploratory data analysis was done on the datasets to improve understanding about the dataset.
Algorithms were constructed to analyze the datasets resulting in a standardized dataset per
patient of 100 slices each having 256x256 pixels. Once the dataset was standardized, two tasks
were executed in parallel. One was the development of web-based software to enable upload,
anonymization and classification of Intracranial Hemorrhage (ICH) CT scans and the other was
the development of a machine learning model for the classification of ICH. The other was the
development of a portal / dashboard for anonymization, transfer and classification of data on
the radiologist side. A prototype that contains features like secure login, secure folder transfer
etc. was built. On the model side, a suite of 3D models that will provide better specificity
compared to the 2D and 2.5D variants was planned. The difficulty in annotation and low sample
volume of data was not enough to aid in convergence of the ML model. This was resolved
through a dataset release by RSNA that helped in building models. Another issue was the lack of
adequate GPU infrastructure (specifically GPU RAM) for building a 3D model. This was resolved
through computer resources provided by CSIR-CEERI. Since the issues of data and infrastructure
look to be temporarily closed, 2D models to classify intracranial hemorrhages were built with up
to 94.6% accuracy.

47