Page 65 - Annual report 2021-22

Annual Report 2021-22 |

variant peptides which are possible candidate proteoforms in different MS datasets of different brain
samples have been identified. These candidate proteoforms can be used as a reference map in brain
proteomics studies.

Detecting and quantifying Intracranial Haemorrhage from Brain Computed Tomography scans using
Deep Learning methods
48
Intracranial Haemorrhage (ICH) is a disabling and often fatal brain condition caused by presence of
blood outside blood vessels within the cranium. ICH is primarily caused due to trauma to the head and
is a time-sensitive condition requiring urgent surgical intervention in order to avoid permanent
disability. The primary way for investigating the presence of a haemorrhage is a Computed
Tomography (CT) scan and requires inputs from radiologists who have a shortage of time and trained
manpower. Therefore, to help radiologists improve the turn-around-time for radiological reports,
Debasis Dash has built AI based solutions for the detection (or classification; 95.6% accuracy),
characterization (or subtype classification; 93.4% accuracy), and quantification (or segmentation;
97.8% accuracy; ~0.8 IOU) of ICH. For the future, a platform for the dissemination of these algorithms
in a remote setting while maintaining privacy of individuals by avoiding transfer of medical data is
planned. The group also aims to develop software to improve the process of generation of training
data for such models in order to improve the Medical AI ecosystem in the country and beyond.

Detecting Interstitial Lung Diseases from Chest Computed Tomography scans using Deep Learning
methods

Interstitial Lung Diseases (ILDs) are a group of heterogeneous disorders resulting from the damage to
the lung parenchyma by inflammation and fibrosis. It leads to impaired alveolar gas exchange which
results in reduced lung function. ILDs are caused by diverse exogenous (environmental and
occupational exposure) and endogenous (autoimmune diseases) factors. The ILD project aims to
develop deep-learning-based methods for the classification and quantification of interstitial lung
diseases (ILDs) using High resolution computed tomography (HRCT) images. Currently, a deep-
learning-based method (Jaccard score of 0.94) for the segmentation of lung lobes from the HRCT slice
using publicly available annotated database of ILD CT scans from the University Hospital of Geneva
has been developed. Additionally, methods for the texture-based classification of ILDs which are
inspired from benchmarked fully connected and object detection computer vision models are being
developed. In collaboration with AIIMS (New Delhi) acquisition and annotation of HRCT data for
training and validation of the developed methods has been initiated.

“If I have seen further, it is by standing on the shoulders of Giants.”
― Isaac Newton