For setting up environment, please read SETUP.md

Polyp Localization and Detection

This work was presented during EEML Summer School, 2019. Find the attached poster here.

For our localization and detection task, we used ColonCVC dataset and ETIS-Larib dataset.

1. Polyp Localization

The task here is to train Fully Convolutional Network (FCN-8s) to create segmentation masks for the polyps and then draw a bounding box around it. To train FCN-8s, run main.py. To evaluate/test the models, run predict_masks.py

2. Polyp Detection

The task here is to train object detection network. We used SSD (Single Shot Multibox detectot to evaluate our results). Code is available for both Faster R-CNN and SSD.

To train faster R-CNN, run the file keras_train_frcnn.py To test faster R-CNN, run the file keras_test_frcnn.py

To run SSD, migrate to directory SSD-keras, and follow README.md for instructions

How to run main.py / predict_masks.py / keras_train_frcnn.py / keras_test_frcnn.py

usage: main.py [-h] [-d DATASET] [--data-dirpath DATA_DIRPATH]
               [--n-workers N_WORKERS] [--gpu GPU] [-rs RANDOM_SEED]
               [-a ARCHITECTURE] [-l LOSS] [-b BATCH_SIZE]
               [-e EPOCHS]  [-lr LEARNING_RATE] [-opt OPTIM] [-m MODEL_NAME]
               [-r RESUME] [-tf TF_LOGS] [-wd WEIGHT_DECAY]
               [-dp DROPOUT]

optional arguments:
  -h, --help            show this help message and exit
  -d DATASET, --dataset DATASET
                        dataset, {'polyps'} (default: polyps)
  --data-dirpath DATA_DIRPATH
                        directory for storing downloaded data (default: data/)
  --n-workers N_WORKERS
                        how many threads to use for I/O (default: 4)
  --gpu GPU             ID of the GPU to train on (or '' to train on CPU)
                        (default: 0)
  -rs RANDOM_SEED, --random-seed RANDOM_SEED
                        random seed for training (default: 1)
  -a ARCHITECTURE, --architecture ARCHITECTURE
                        network architecture name, {'fcn8s1'}
                        (default: fcn8s1)
  -l LOSS, --loss LOSS
                        loss function, {'cse'} (default: cse, mse, dse)
  -b BATCH_SIZE, --batch-size BATCH_SIZE
                        input batch size for training (default: 50)
  -e EPOCHS, --epochs EPOCHS
                        number of epochs (default: 100)
  -lr LEARNING_RATE, --learning-rate LEARNING_RATE
                        initial learning rate (default: 0.0001)
  -opt OPTIM, --optim OPTIM
                        optimizer, {'adam', 'sgd', 'adagrad', 'rms_prop'} (default: adam)
  -m MODEL_NAME, --model_name MODEL_NAME
                        name of the model (default: 'fcn8s_1')
  -r RESUME, --resume RESUME
                        flag to resume training (default: 'y')
  -tf TF_LOGS, --tf_logs TF_LOGS
                        path for tensorboaard loggging (default: 'tf_logs')
  -wd WEIGHT_DECAY, --weight_decay WEIGHT_DECAY
                        weight decay (default: 0)
  -dp DROPOUT, --dropout DROPOUT
                        dropout (default: 0)

Sample Command

python3 main.py -d polyps -a fcn8s -b 100  -e 20

3. Misc

Code is available for both Faster R-CNN and SSD. But I used SSD for the task of detecting polyps.

All the results and models can be downloaded from here. No pretrained models were used. Everything is trained from scratch.

4. Results

Results on Validation set for Single Shot Multi Box Detector

Succesful Cases (Benchmark Dataset CVC Colon)

Failure Cases (Benchmark Dataset CVC Colon)