AI Chest X-ray Diagnosis API

TL;DR

This project provides 18-pathology diagnosis from chest x-rays using TorchXRayVision. It is deployed on AWS Lambda and must be accessed via a AWS SDK, as AWS API Gateway is not feasible due to exceeding the 30-second timeout limit when waking from a cold start.

Features

• Diagnosis for 18 chest X-ray pathologies, including Atelectasis, Cardiomegaly, Consolidation, Edema, Effusion, Emphysema, Enlarged Cardiomediastinum, Fibrosis, Fracture, Hernia, Infiltration, Lung Lesion, Lung Opacity, Mass, Nodule, Pleural Thickening, Pneumonia, and Pneumothorax.
• Detection of out-of-distribution images using an auto-encoder to prevent predictions on images that are different from the training data.
• Grad-CAM visualisation to highlight regions of interest in X-ray images

Technologies

• TorchXRayVision: For chest X-ray analysis
• GitHub Actions
• Docker
• Terraform
• AWS Lambda

How to use

• Function name: ai-chest-xray-diagnosis-api

• Payload

  • base64Img (string, required): The base64-encoded string of the chest x-ray image.
  • reconstructionThreshold (float, optional): The threshold for the reconstruction error to determine if the image is in-distribution. Default is 8000.
  • ssimThreshold (float, optional): The Structural Similarity Index (SSIM) threshold to determine if the image is in-distribution. Default is 0.62.
  • gradCamThreshold (float, optional): The threshold for pathology scores above which Grad-CAM images are generated. Default is 0.44.

• Example Request:

{
  "base64Img": "iVBORw0KGgoAAAANSUhEUgAAB4AAAAL...",
  "reconstructionThreshold": 8000,
  "ssimThreshold": 0.62,
  "gradCamThreshold": 0.44
}

• Example Response (200 OK)

{
  "reconstructionError": 3334.66796875,
  "ssim": 0.7661488930884929,
  "inDistribution": true,
  "prediction": {
    "Atelectasis": 0.5357859134674072,
    "Consolidation": 0.14535437524318695,
    "Infiltration": 0.5115969181060791,
    "Pneumothorax": 0.08925247937440872,
    "Edema": 0.0008359009516425431,
    "Emphysema": 0.5003246068954468,
    "Fibrosis": 0.5113949179649353,
    "Effusion": 0.11136345565319061,
    "Pneumonia": 0.0503874197602272,
    "Pleural_Thickening": 0.5349103212356567,
    "Cardiomegaly": 0.4702714681625366,
    "Nodule": 0.5080699920654297,
    "Mass": 0.5646909475326538,
    "Hernia": 0.995814323425293,
    "Lung Lesion": 0.0018308708677068353,
    "Fracture": 0.5114134550094604,
    "Lung Opacity": 0.2932303249835968,
    "Enlarged Cardiomediastinum": 0.10778136551380157
  },
  "gradCam": {
    "Atelectasis": "iVBORw0KGg...",
    "Infiltration": "iVBORw0...",
    "Emphysema": "iVBORw0KGg...",
    "Fibrosis": "iVBORw0KAAN...",
    "Pleural_Thickening": "iVBORw0K...",
    "Cardiomegaly": "iVBORw...",
    "Nodule": "iVBORw0KGg...",
    "Mass": "iVBORw0KGgoA...",
    "Hernia": "iVBORw0KG...",
    "Fracture": "iVBORw0KGgoA...",
  }
}