tensorflow

Google Object Detection API

We used Google's Object Detection API to train various object detection models: SSD Mobilenet v2, Faster RCNN Inception v2, Faster RCNN Resnet50 and Faster RCNN Resnet101.

How to set up Google's Object Detection API

The detailed instructions can be found on the official website, but here is the overview:

Step 1: Install the required packages

Simply run pip install -r requirements.txt from the root directory.

Step 2: Instll COCO API

git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
make
cp -r pycocotools <root_dir>/tensorflow/tf-models/research/

Step 3: Compile protobuf

# from <root_dir>/tensorflow/tf-models/research/
protoc object_detection/protos/*.proto --python_out=.

Step 4: Set PYTHONPATH

Follow the instructions on the main README.md.

How to train

Step 1: Convert DICOMs to Tensorflow Records

Use create_pneumonia_tf_record.py to convert DICOM images to Tensorflow Records.

python create_pneumonia_tf_record.py --dicom_dir /dir/to/DICOMs --eval_perct 0.2 --label_file /path/to/label/file --output_dir data/records

Step 2: Download pretrained models

You can download pretrained models from Google's Object Detection API Model Zoo. We used the following:

• SSD Mobilenet v2
• Faster RCNN Inception v2
• Faster RCNN Resnet50
• Faster RCNN Resnet101

The sample config files expect the pretrained models to be in the model file. For example, for SSD Mobilenet v2, place the pretrained model here.

Step 3: Train

Use Google's Object Detection API to train the models. The config files are here.

# train SSD Mobilenet v2
python tf-models/research/object_detection/model_main.py \
    --pipeline_config_paath models/ssd_mobilenet_v2/ssd_mobilenet_v2_pneumonia.config \
    --model_dir /path/to/output

Step 4: Export the trained model for inference

Once you're done with training, you need to export the model so that you can use it for inference. You can also override some of the model config values such as the max suppression confidence threshold and IOU threshold.

# without config override
python tf-models/research/object_detection/export_inference_graph.py \
    --pipeline_config_path /path/to/trained/pipeline.config \
    --trained_checkpoint_prefix /path/to/trained/model.ckpt-<step_number> \
    --output_directory /output/dir

# with config override for faster rcnn
python tf-models/research/object_detection/export_inference_graph.py \
    --pipeline_config_path /path/to/trained/pipeline.config \
    --trained_checkpoint_prefix /path/to/trained/model.ckpt-<step_number> \
    --output_directory /output/dir \
    --config_override "
    model {
        faster_rcnn {
            second_stage_post_processing {
                batch_non_max_suppression {
                    score_threshold: 0.3
                    iou_threshold: 0.1
                    max_detection_per_class: 10
                    max_total_detection: 10
                }
            }
        }
    }"

Step 5: Evaluate

Use model_eval.py to generate evaluation images with detected boxes to see how your model performs.

python model_eval.py \
    --label_map data/pneumonia_label_map.pbtxt \
    --inference_graph /path/to/exported/frozen_inference_graph.pb \
    --input_tfrecord_pattern "data/records/pneumonia_eval.record-?????-of-00010" \
    --output_images_dir /path/to/output

Step 6: Test

Use model_infer.py to test the model on new DICOM images and generate a submission file.

python model_infer.py \
    /path/to/exported/frozen_inference_graph.pb \
    data/pneumonia_label_map.pbtxt \
    /path/to/test/dicoms \
    /path/to/submission/file \
    /path/to/output/image/dir