Detectos® - Chest X-Ray Screening Platform

[2018-07-31] A radiologists buddy for health check-up chest X-Ray screening

What is Detectos® ? #

• It is an AI model that BridgeAsia team has created for health check-up X-Ray Screening
• It gives a score between 0 - 100 for each X-Ray image - the higher number = higher abnormal possibility

Pain Points #

• The client hospital has around 100K-200K health check-up X-Ray images per year
• The abnormal prevalence around 5%, so finding the problematic images are harder
• The radiologists are in high workload and tend to have less efficiency overtime

Solution #

• AI read through the images first to flag high abnormal probability images
• The radiologists focus on the flagged images first
• The radiologists still need to read all images, but has a better choice to manage their time and energy

Detectos® Performance #

• We set the sensitivity target to 93% (The radiologist sensitivity from this paper is 92.6%)
• At 93% sensitivity, the model can save radiologists effort by 34.50%

System Design #

There are three main components:

• Abnormality Score Manager = AI Model + DICOM SCP
• Detectos Perpetual Learning = Labeling Tool, DataSet Builder
• Detectos Model Trainer = Python Script to train AI model by command-line

ASM - Design

Labeling System UI screen shot

Software Stack #

Abnormality Score Manager #

• Python, Flask, fastai, PyTorch, pydicom, pynetdicom
• rabbitmq

Requirement.txt

pydicom==1.3.0
pynetdicom==1.4.1
celery==4.4.2
Pillow==6.0.0
numpy==1.16.1
torchvision==0.2.2.post3
torch==1.0.0
fastai==1.0.51
mysql-connector-python==8.0.17
scikit-learn==0.20.3
asyncio==3.4.3
flower==0.9.3
pyarmor==5.5.6
PyJWT==1.7.1
cryptography==2.7
Flask==1.1.1
Flask-RESTful==0.3.7
Flask-JWT-Extended==3.22.0
waitress==1.3.1
pyarmor==5.5.6
suds-community==0.8.4

Detectos Perpetual Learning #

• PHP7.2, Phalcon3
• cornerstone js - web based medical imaging platform