1 of 1

Solution

We developed Deseray, our novel patch-based fully convolutional encoder-decoder neural network. We trained Deseray on a publicly available dataset of over 12,000 X-rays and on our curated dataset of over 1,000 X-rays from the Radiology Department of The Medical City (TMC) in Pasig City, Philippines.

Problem Statement

Detecting pneumothorax on chest X-rays is challenging for radiologists because its sole visual indicator on the X-ray is often just a thin, displaced pleural line.

Features

• clinical support for radiologists: Deseray provides accurate detection and precise segmentation on chest X-rays, with specificity of 99%, sensitivity of 84%, and mean DSC of 91% on the local TMC test dataset
• efficiency: Deseray has an average inference time of 0.3184 seconds per image
• human-in-the loop AI: Radiologists at TMC are in the loop of improving Deseray through our in-house annotation tool
• cloud-based web app: interested users can inquire and get access at www.deseray.io

DESERAY: Detection and Segmentation of

Pneumothorax on Chest X-rays [ www.deseray.io ]

J.I. Dumbrique, R. Hernandez, J.M. Cruz, M.R. Pagdanganan, and P.C. Naval, Jr.

chest X-ray of patient diagnosed with pneumothorax

ground-truth mask annotated by radiologists at TMC

Deseray’s output�DSC = 93.7%�IoU = 88.2%