Your Name: Josip Vrdoljak
Name of your Device: Pneumonia detection algorithm
Intended Use Statement: Assisting clinitians in pneumonia detection
Indications for Use: Screening for pneumonia in patients aged 20-80 years, at Infectious disease and Respiratory disease emergency rooms
Device Limitations: ROC-AUC of 0.77. Worse sensitivity in the preseance of infiltrations.
Clinical Impact of Performance: Faster screening times, can help doctors in the triage of patients
<< Insert Algorithm Flowchart >>
DICOM Checking Steps: - check wether the image is AP or PA
Preprocessing Steps: - resized images and performed image augmentation
CNN Architecture: - VGG 16 first 16 layers + 1 final dense layer with a sigmoid activation function
Parameters:
- ImageDataGenerator function for Keras was used for augmentation
- Batch size: 32
- Optimizer learning rate: 1e-4
- Layers of pre-existing architecture that were frozen: 16
- Layers of pre-existing architecture that were fine-tuned: 0
- Layers added to pre-existing architecture: 1
<< Insert algorithm training performance visualization >>
<< Insert P-R curve >>
Final Threshold and Explanation: We picked 0.64 as an optimal threshold based on the P-R curve
(For the below, include visualizations as they are useful and relevant)
Description of Training Dataset: 2290 images extracted and augmented from the NIH chest x-ray dataset
Description of Validation Dataset: 1390 images extracted and augmented from the NIH chest x-ray dataset Both datasets were confirmed to have the same distributions of the target variable (pneumonia)
Was established via radiologist findings.
Patient Population Description for FDA Validation Dataset: Patients that are suspected to have pneumonia
Ground Truth Acquisition Methodology: radiologist finding
Algorithm Performance Standard: AUC of 0.77