Photo by Amanda Dalbjörn on Unsplash
Samsung (OCT) Disease Prediction is a utility that facilitates the diagnosis of retinal disease from optical coherence tomography images. This project utilizes supervised and semi-supervised techniques to generate models trained at multi-class disease classification and exposes model inference through a collection of REST-ful API endpoints. The full project is deployable as a stand-alone, containerized application or on popular cloud-computing environments.
The OCT (Optical Coherence Tomography) is an imaging method used to diagnose the patient’s retinal health into four categories:
- Normal
- CNV
- DME
- DRUSEN
- create a baseline: reproduce the SOTA classification accuracy by training a Deep CNN (Supervised Learning) We use Inception_V3 model for this part.
- improve the classification accuracy over the baseline in part 1 using the self-supervised method described in simclr paper.
- Learn the Unsupervised Features based on the Self-Supervision technique
- Train a classification head on top of the learned Unsupervised Features to further improve the accuracy upon the established baseline.
-
This tool can expedite the diagnosis of treatable diseases that can lead to blindness.
-
By doing so, medication can be prescribed in time for the patients and prevent them from becoming blind.
-
Additionally, the tool has the potential to be generalized in other applications of biomedical imaging including x-ray, MRI, and computer tomography)
To give some context, 30 million OCT scans are produced each year. Nearly 11 million people in the US alone will suffer from 1 of these diseases.
- The outcome is a support tool that provides retinal specialists with accurate and timely diagnosis of key pathology in OCT images.
- The OCT images can be used both for populations without proper access, and for those in overpopulated areas in order to reduce patient burden by using automated triage systems.
[Mendeley] (https://data.mendeley.com/datasets/rscbjbr9sj/2)
Model inference for supervised and semi-supervised learning methods are exposed through API endpoints using FastAPI. For convenience, the utility comes included with trained models -- an InceptionV3 neural network for supervised image classification and a ResNet-50 neural network utilizing the SimCLRv2 methodology -- available in the src/classifier/model directory. Furthermore, the repository includes a set of sample instances that were unseen from either model, all available in the src/corpus directory.
- Python 3.9+
- Docker
The included Dockerfile allows for building a minimal container environment to run and test the API.
To build a docker image from the project, run the following command from the root project directory:
docker build -t samsung-oct .
To create a new container, run the following command:
docker run -d --name <CONTAINER_NAME> -p 8000:8000 samsung-oct
NOTE The API may not be immediately available after running a container since the saved models need to be loaded.
The root endpoint to test whether the API service is running and functional.
GET /None
{
"service": "inference"
"status": "online"
}
An endpoint to return a classification report from all instances in the corpus. Either the inceptionv3 or simclrv2 models can be used.
GET /classification_report/?model_name={{model_name}}| Parameter | Type | Description |
|---|---|---|
model_name |
string |
Required. Either inceptionv3 or simclr to generate a classification report. |
{
"classification_report": {
"0": {
"precision": 0.6666666666666666,
"recall": 1.0,
"f1-score": 0.8,
"support": 4
},
"1": {
"precision": 0.375,
"recall": 0.75,
"f1-score": 0.5,
"support": 4
},
"2": {
"precision": 1.0,
"recall": 0.5,
"f1-score": 0.6666666666666666,
"support": 4
},
"3": {
"precision": 0.0,
"recall": 0.0,
"f1-score": 0.0,
"support": 4
},
"accuracy": 0.5625,
"macro avg": {
"precision": 0.5104166666666666,
"recall": 0.5625,
"f1-score": 0.4916666666666667,
"support": 16
},
"weighted avg": {
"precision": 0.5104166666666666,
"recall": 0.5625,
"f1-score": 0.4916666666666667,
"support": 16
}
}
}
An endpoint to return predictions for varying number of instances from the corpus.
NOTE: When stratification is used, sampling across each class will be conducted. The API will return the max number of instances in teach class if num_samples exceeds the total available instances in a given class in the corpus. When stratification is not used, random sampling across classes will be conducted.
GET /corpus_predict/?model_name={{model_name}}&num_samples={{num_samples}}&stratify={{stratify}}| Parameter | Type | Description |
|---|---|---|
model_name |
string |
Required. Either inceptionv3 or simclr for running inference. Defaults to inceptionv3. |
num_samples |
int |
The number of samples to run inference against. Defaults to 1. |
stratify |
bool |
Indicator for whether sampling should be stratified across all classes. Defaults to False. |
{
"predictions": [
{
"index": 0,
"instance": "DRUSEN-224974-4.jpeg",
"label": "DRUSEN",
"prediction": 0
}
]
}
Unit tests and API tests can be run after installing requirements:
python -m pytest src/tests/unit/test_utilities.py
python -m pytest src/tests/unit/test_utilities.py