Lesion Detection and Analysis in CT Scan

Project Title

Lesion Detection and Analysis in CT Scan

Project Area of Specialization Artificial IntelligenceProject Summary

The intended system aims at making Report-Writing easier and efficient for Radiologists by automatically detecting and analyzing lesions (abnormalities) in a CT Scan.

Traditionally, a radiologist takes on average 1 to 2 hours to view a CT Scan, go through its slices, drop scales manually and adjust intensities to get an authentic picture of the patient's condition. Therefore, it takes him over an hour to report a ct scan. Also, radiologists have to report x-rays, ultrasounds, MRI etc which makes their job burdensome. Eventually, it effects the patient who suffers huge delays in getting report of their scan.

Our end product is a desktop application to be used by radiologists. It would take a CT Scan as an input and display a 3D view of the scan with lesions highlighted and labeled. Labeling would include all the essential details about the lesion that a Radiologist is concerned with. It includes but is not limited to:

1. Shape

2. Surface Type

3. Density

4. Dimensions

5. Location

6. Organs involved

The intended system will generate a crisp report to only be cross-checked by the radiologist before approving it for use by the Physician concerned.

Project Objectives

The main objectives of the project are to

1. Accurately and Precisely detect and segment a lesion. It includes drawing a mask over the lesion.

2. Ascertain shape of the lesion.

3. Determine surface type

4. Assess whether it has involved any organs and up to what extent.

5. Denote all the findings in a human-readable language.

The end goal of the project is to minimize the shortage of Radiologists by making their job easier and faster.

Project Implementation Method

The first part of the project involves detecting lesions. For this purpose, we have trained a deep learning model, based on the U-Net architecture of a convolutional neural network. The model was trained on the "DeepLesion" dataset available on the National Institute of Health America's public repository. It includes 32,000 annotated lesions spread across 130,000 CT image slices.

The model generates masks over the lesion with 65% mean_IoU so far. we intend to increase it to 90% but for that we need computational resources and more data.

The next step would be to design rule-based and intelligent agents that would assess the shape using density based clustering and suface type by using "tangantial deviation".

The last part would be converting the information into meaningful text. It woud initially be hard coded but eventually it would use NLP to provide descriptive reports.

Benefits of the Project

Getting CT Scaned takes less than 30 minutes but getting the report requires 2 to 5 working days depending on the service provider. Very often, due to this delay in reporting, the condition of the patient worsens to a point of no return. Our projectr aims to save lives by providing report of the scan instantly.

In PIMS for example, it takes on minimum, 5 days to process a ct scan given the shortage of doctors in genral and radiologists in particular.

In 2018, 36,000 CT scans were performed in PIMS out of which 25,000 were reported in an emergency so they couldn't wait for traditional reporting and hence they were assessed by Medical Officer on call. Such alternatives amount to erroneous diagnoses and loss of precious lives. For the rest of the scans, the delay in reporting causes deteriorated health condition.

Over an hour on a single ct scan with scarce skilled human resources causes an overhead that keeps piling up in public sector hospitals. Our system will ensure runtime reporting which can be cross-checked by a radiologist taking less than 5 minutes on a scan. As a result, the patient will get their report right when they come out of laboratory.

It would

1. Save Lives

2. Save resources

3. Lead to better utilization of costly CT Machines

Technical Details of Final Deliverable

Our Final deliverable includes

1. Front-end implemented in Electron JS

2. The desktop application with report generation feature implemented

Final Deliverable of the Project Software SystemCore Industry MedicalOther Industries IT Core Technology Artificial Intelligence(AI)Other Technologies Artificial Intelligence(AI)Sustainable Development Goals Good Health and Well-Being for People, Industry, Innovation and InfrastructureRequired Resources

Item Name	Type	No. of Units	Per Unit Cost (in Rs)	Total (in Rs)
			Total in (Rs)	80000
GPU	Equipment	1	70000	70000
Data Annotation	Miscellaneous	1	10000	10000