Reconstructing Sinus Anatomy from Endoscopic Video -- Towards a Radiation-free Approach for Quantitative Longitudinal Assessment

Left to right: original video, reconstruction fly-through, depth rendering, monocular depth estimate, monocular depth uncertainty estimate

This codebase implements the method described in the paper:

Reconstructing Sinus Anatomy from Endoscopic Video -- Towards a Radiation-free Approach for Quantitative Longitudinal Assessment

Xingtong Liu, Maia Stiber, Jindan Huang, Masaru Ishii, Gregory D. Hager, Russell H. Taylor, Mathias Unberath

In the 23rd International Conference on Medical Image Computing and Computer-Assisted Intervention (MICCAI 2020)

Please contact Xingtong Liu ([email protected]) or Mathias Unberath ([email protected]) if you have any questions.

We kindly ask you to cite this paper if the code is used in your own work.

@InProceedings{liu2020reconstructing,
author="Liu, Xingtong
and Stiber, Maia
and Huang, Jindan
and Ishii, Masaru
and Hager, Gregory D.
and Taylor, Russell H.
and Unberath, Mathias",
editor="Martel, Anne L.
and Abolmaesumi, Purang
and Stoyanov, Danail
and Mateus, Diana
and Zuluaga, Maria A.
and Zhou, S. Kevin
and Racoceanu, Daniel
and Joskowicz, Leo",
title="Reconstructing Sinus Anatomy from Endoscopic Video -- Towards a Radiation-Free Approach for Quantitative Longitudinal Assessment",
booktitle="Medical Image Computing and Computer Assisted Intervention -- MICCAI 2020",
year="2020",
publisher="Springer International Publishing",
address="Cham",
pages="3--13",
isbn="978-3-030-59716-0"
}

Instructions

1. Install all necessary python packages:

torch, torchvision, opencv-python (<= 3.4.2.16), opencv-contrib-python (<= 3.4.2.16), numpy, tqdm, pathlib, torchsummary, tensorboardX, albumentations, argparse, pickle, plyfile, pyyaml, datetime, random, shutil, matplotlib, tensorflow, autolab_core, autolab_perception, meshrender, h5py, trimesh, scikit-image, sqlite3

Replace the python scripts in the meshrender with the ones in this repo to support texture rendering.

2. Generate SfM results from videos using Structure from Motion (SfM), e.g. COLMAP.

3. Convert SfM results to the format similar to this storage. Color images with the format {:08d}.jpg are extracted from the video sequence. camer_intrinsics_per_view stores fx, fy, cx, cy element of the estimated camera intrinsics. In this example, since all images are from the same video sequence, we assume the intrinsics are the same for all frames. motion.yaml stores the estimated poses of the camera coordinate system w.r.t. the world coordinate system. selected_indexes stores all frame indexes of the video sequence. structure.ply stores the estimated sparse 3D reconstruction from SfM. undistorted_mask.bmp is a binary mask used to mask out blank regions of the video frames. view_indexes_per_point stores the indexes of the frames that each point in the sparse reconstruction gets triangulated with. The views per point are separated by -1 and the order of the points is the same as that in structure.ply. We smooth out the point visibility information in the script to make the global scale recovery more stable and obtain more sparse points per frame for training. The point visibility smoothness is controled by parameter visibility_overlap. visible_view_indexes stores the original frame indexes of the registered views where valid camera poses are successfully estimated by SfM. To convert SfM results from COLMAP to the required format, one example for using colmap_model_converter.py is:

/path/to/python /path/to/colmap_model_converter.py --colmap_exe_path "/path/to/COLMAP.bat" --sequence_root "/path/to/video/sequence"

4. Train a feature descriptor model for pair-wise feature matching in SfM. One example for using train_descriptor.py is:

/path/to/python /path/to/train_descriptor.py --adjacent_range 1 50 --image_downsampling 4.0 --network_downsampling 64 --input_size 256 320 --batch_size 4 --num_workers 4 --lr_range 1.0e-4 1.0e-3 --inlier_percentage 0.9 --display_interval 50 --validation_interval 2 --training_patient_id 1 --load_intermediate_data --num_epoch 60 --num_iter 1000 --heatmap_sigma 5.0 --visibility_overlap 20 --display_architecture --data_root "/path/to/training/data/root" --sampling_size 10 --log_root "/path/to/log/root" --feature_length 128 --filter_growth_rate 10 --matching_scale 20.0 --matching_threshold 0.9 --rr_weight 1.0 --cross_check_distance 3.0 --precompute_root "/path/to/precompute/root"

--load_trained_model --trained_model_path "/path/to/trained/model" can be added to continue previous training.

5. Run dense_feature_matching.ply with the trained descriptor model to generate pair-wise feature matches for SfM running. One example is:

/path/to/python /path/to/dense_feature_matching.py --image_downsampling 4.0 --network_downsampling 64 --input_size 256 320 --batch_size 1 --num_workers 1  --load_intermediate_data --data_root "/path/to/data/root" --sequence_root "/path/to/sequence/data" --trained_model_path "/path/to/trained/descriptor/model" --precompute_root "/path/to/precompute/root" --feature_length 128 --filter_growth_rate 10 --max_feature_detection 3000 --cross_check_distance 3.0 --patient_id 1 --gpu_id 0 --temporal_range 30 --test_keypoint_num 200 --residual_threshold 5.0 --octave_layers 8 --contrast_threshold 5e-5 --edge_threshold 100 --sigma 1.1 --skip_interval 5 --min_inlier_ratio 0.2 --hysterisis_factor 0.7     

6. Run colmap_database_creation.py to convert the generated feature matches in HDF5 format to SQLite format, named database.db, that is compatible with COLMAP. One example is:

/path/to/python /path/to/colmap_database_creation.py --sequence_root "/path/to/video/sequence"

7. Run colmap_sparse_reconstruction.py to run mapper in COLMAP for bundle adjustment to generate sparse reconstruction and camera trajectory. One usage example is:

/path/to/python /path/to/colmap_sparse_reconstruction.py --colmap_exe_path "/path/to/COLMAP.bat" --sequence_root "/path/to/video/sequence"

8. Run colmap_model_converter.py again to convert the SfM result from COLMAP with the trained feature descriptor to the required format for training a depth estimation model. See step 3 for one example.

9. Train a depth estimation model with the data generated in step 8. The previous trained descriptor model can be used to calculate an optional appearance consistency loss. One example is:

/path/to/python /path/to/train_depth_estimation.py --adjacent_range 5 30 --image_downsampling 4.0 --network_downsampling 64 --input_size 256 320 --batch_size 4 --num_workers 4 --slp_weight 1.0 --dcl_weight 0.5 --sfl_weight 2.0 --dl_weight 0.05 --lr_range 1.0e-4 1.0e-3 --inlier_percentage 0.9 --display_interval 20 --visible_interval 5 --save_interval 1 --training_patient_id 1 --num_epoch 40 --num_iter 1000 --display_architecture --load_intermediate_data --data_root "/path/to/training/data/root" --log_root "/path/to/log/root" --precompute_root "/path/to/precompute/root" --descriptor_model_path "/path/to/trained/descriptor/model"

--load_trained_model --trained_model_path "/path/to/trained/model" can be added to continue previous training.

10. Run fusion_data_generation.py to generate depth estimates for model fusion later. One example is:

/path/to/python /path/to/fusion_data_generation.py --image_downsampling 4.0 --network_downsampling 64 --input_size 256 320 --batch_size 1 --num_workers 1 --visible_interval 5 --inlier_percentage 0.9 --load_intermediate_data --trained_model_path "/path/to/trained/model" --data_root "/path/to/data/root" --sequence_root "/path/to/sequence/root" --patient_id 1 --precompute_root "/path/to/precompute/root"

11. Run surface_reconstruction.py to generate the textured watertight surface reconstruction. A fly-through video will also be generated for sanity check. One example is:

/path/to/python /path/to/surface_reconstruction.py --data_root "/path/to/data/root" --visualize_fused_model --trunc_margin_multiplier 10.0 --sequence_root "/path/to/sequence/root" --patient_id 1 --max_voxel_count 64e6

Pre-trained models

The pre-trained weights for dense descriptor network and depth estimation network are provided here.

Related Projects

• SAGE: SLAM with Appearance and Geometry Prior for Endoscopy (ICRA 2022)

• Neighborhood Normalization for Robust Geometric Feature Learning (CVPR 2021)

• Extremely Dense Point Correspondences using a Learned Feature Descriptor (CVPR 2020)

• Dense Depth Estimation in Monocular Endoscopy with Self-supervised Learning Methods (TMI)