Page 49 - Biennial Report 2018-20 Jun 2021

DEEP LEARNING FOR CLASSIFICATION AND QUANTIFICATION OF INTERSTITIAL
LUNG DISEASES.

Interstitial lung disease is a group of disorders which causes progressive damage to lungs. It is an
umbrella terminology for more than 200 different abnormalities that hinders proper respiration
by scarring the lung tissue. Some of these progressive disorders such as Idiopathic Pulmonary
Fibrosis (IPF) is an irreversible degradation of lung function which eventually leads to human
body collapse in less than 3 years after detection. There is no such permanent restoration
solution available for IPF, other than bringing down the intensity of symptoms related to the
same. Open Lung Biopsy being a prime examiner of disorders rather carries a risk of serious
complications. High Resolution Computed Tomography (HRCT) is preferred over other invasive
tests. HRCT characterizes ILDs with clarity of textures and crucial details resulting in an accurate
initial diagnostic. Due to the dearth of radiologists and field experts, Machine Learning/Deep
Learning is therefore an appropriate solution. Lung functioning adapts the environment and so
textures vary with the location; hence local data is most suitable for training the Machine
Learning model/Deep Learning.

As per the inclusion and exclusion criteria outlined during the study design, patients were
recruited, and data collection is currently under progress. Ethical permissions and other
necessary clearances were obtained. These were done in collaboration with AIIMS, New Delhi.
Furthermore, labeling and annotation of the CT scans were performed, and an algorithm is being
developed for texture-based classification of data and its quantification, by CSIR-CEERI. For this
phase, apart from the AIIMS dataset, an access to the publicly available multimedia database of
interstitial lung diseases was also obtained from the University Hospital of Geneva. The algorithm
is being developed with the help of machine learning techniques such as recurrent neural
network and generative adversarial network to create a predictive model for texture-based
classification and quantification of data and to account for the limitations of the dataset.

Industrial partner ‘Predible Health’ was identified, and initial talks have been held with regard to
testing the machine learning models, being developed, for texture-based classification and
quantification of CT scan data.

CHRONIC RESPIRATORY DISEASE INNOVATION AND SOLUTION PROGRAM

Chronic Obstructive Pulmonary Disease (COPD) is a condition of the lung caused by genetic
factors, behavioral practices like smoking and environmental factors like pollution. In rural
households, cooking in poorly ventilated rooms can lead to COPD. The patient suffers from
progressive symptoms like shortness of breath and poor airflow. Chronic inflammation of the
lung epithelium accompanied by the constriction of airways are part of the changes that lead to
COPD.

Under this project, moving from molecular mechanism to solutions, the objective is to evaluate
the therapeutic potential of PQQ, a microbial derived compound in patients with IPF and COPD.
The project also focuses on the role of phosphatase inactive INPP4A in reducing allergic airway
inflammation. Asthma is also a chronic respiratory disease, marked by acute episodes
of wheezing, coughing, chest tightness, and shortness of breath. The primary difference

48