Page 55 - CSIR-IGIB Annual Report 2020-21
P. 55

RAPID-CT project for CT scan triage  and          segmentation algorithm of lung mask detection
                diagnosis                                         on 100 chest X-rays as test set, from the pool of
                RAPID-CT aims to reduce turnaround time for       1000  X-rays from RSNA dataset. We obtained
                triage of medical images while also prioritization   0.91 as Jaccard similarity index on the validation
                of patients based on their condition in a clinical   set. We validated the lung opacity  detection
                setting automatically.  As a case study,          module on 1012 test X-ray images from RSNA
                Computed Tomography scans of the  Head for        Kaggle dataset. The exactness of object
                detection and diagnosis of Intracranial           detection  is usually  well determined by mAP
                Hemorrhages (ICH) has been selected as an         (Mean Average Precision), for opacity detection
                initial challenge. RAPID-CT also aims to provide   mAP of 0.34 is achieved. CovBaseAI algorithm
                a software for assessment of CT scans in a        was found  to  have  an accuracy  of  87% with
                remote setting. Apart from patient  samples       negative  predictive value of 98% in the
                available publicly from  RSNA  to be  used for    quarantine-center  data   for   Cov-Pneum.
                training, we  collected  patient samples from a   However, sensitivity varied from 0.66 to 0.90
                radiology clinic  as a test set. We then  built   depending  on whether RT-PCR or  the
                modules for anonymization and standardization     radiologist’s  opinion was  set as ground truth.
                of CT, and for collection  and inference of the   Since  the CovBaseAI initiative several new
                models which we then integrated  into a web       datasets for SARS-CoV detection are  publicly
                portal. The models we built for detection of ICH   available which needs to be incorporated in the
                presence and subtypes, and localization of ICH    algorithm. Further, there are plans to work on
                on the CT slices had an accuracy of 95.6% and     comparing the efficacy of CXRs as compared to
                98% for ICH detection respectively. In the future   CT scans for detection of COVID-19 and to build
                we plan to work on the following avenues: (i)     a better version of the model.
                Build better models for ICH detection, diagnosis
                and localization including  patient level         Brain tissue-specific proteoforms
                detection models. (ii) Work on generalizability   The human  brain is a  complex network of
                of models in multiple hospitals in a clinical     structural  and   functional  systems.  Its
                setting, and  (iii) Translate skills and toolsets   complexity  is majorly governed by the
                developed for ICH into other use cases in the     expressed proteins. Alterations or mutations at
                near future.                                      different levels from genome to proteome give
                                                                  rise to various proteoforms. Various studies
                CovBaseAI–COVID 19 detection and diagnosis        have shown  that  proteoforms not only show
                CovBaseAI project aims to develop an AI           distinctive tissue specificity  but also  lead  to
                classifier for detection and diagnosis of COVID-  variability in phenotypic traits. Diversity in
                Pneumonia from Chest X-Rays (CXRs). The AI        peptides that gives rise to different proteoforms
                classifier predicts Covid-Pneumomia and is        is important in several neurological disorders
                composed of an ensemble model consisting of       like Alzheimer’s disease. The proteoforms
                three DL modules feeding to an expert decision    cannot be directly predicted from the genome
                system. The DL modules as part of the ensemble    but using  proteogenomics for integrating
                include   pathology    classification,  lung      genome/transcriptome with proteomics data
                segmentation, and opacity detection models        can reveal deeper insights into the identification
                and are explainable  to the extent of an          and tissue-specific  expression of such novel
                activation  map  output. The expert decision      proteoforms. Thus, a  detailed proteogenomic
                system is a rule-based classification system that   study analyzing proteins across different tissues
                classifies the X-ray into one of three  classes,   of the human brain can provide information on
                namely COVID-unlikely, indeterminate, and         how proteoforms relate to human biology and
                COVID-likely and is fully explainable as well as   disease.   We conducted  deep proteogenomic
                modifiable  as needed. We validated  the          profiling of publicly available proteomics data




               Annual Report 2020-21                                                                       52
   50   51   52   53   54   55   56   57   58   59   60