Dataset Release for "LensID: A CNN-RNN-Based Framework Towards Lens Irregularity Detection in Cataract Surgery Videos"

The current dataset includes video annotations and lens/pupil segmentations being created for the paper: "LensID: A CNN-RNN-Based Framework Towards Lens Irregularity Detection in Cataract Surgery Videos". We use three datasets for this study:

Dataset 1

The first set is a large dataset containing the annotations for the lens implantation phase versus the rest of phases from 100 videos of cataract surgery. since lens implantation is a very short phase (around four seconds) compared to the whole surgery (seven minutes on average), creating a balanced dataset that can cover the entire content of videos from the "Rest" class is quite challenging. Hence, we propose a video clip generator that can provide diverse training sequences for the recurrent neural network by employing stochastic functions. At first, 12 three-second video clips with overlapping frames are extracted from the implantation phase of each cataract surgery video. Besides, the video segments before and after the implantation phase are divided into eight and four video clips, respectively (these clips have different lengths depending on the length of the input video). Accordingly, we have a balanced dataset containing 2040 video clips from 85 videos for training and 360 video clips from the other 15 videos for testing. For each training example, the video generator uses a stochastic variable to randomly select a three-second clip from the input clip. We divide this clip into N sub-clips, and N stochastic variables are used to randomly select one frame per sub-clip (in our experiments, N is set to five to reduce computational complexity and avoid network overfitting).

Examples for Implantation Phase:

Examples for the "Rest" class:

Dataset 2

The second set is a dataset containing the lens segmentation of 401 frames from 27 videos (292 images from 21 videos for training, and 109 images from six videos for testing).

Examples for lens segmentation:

Dataset 3

The third set is a dataset containing the pupil segmentation of 189 frames from 16 videos (141 frames from 13 videos for training, and 48 frames from three videos for testing).

Examples for pupil segmentation:


Disclaimer

The datasets are exclusively provided for scientific research purposes and as such cannot be used commercially or for any other purpose. If any other purpose is intended, you may directly contact the originator of the dataset, Prof. Yosuf El-Shabrawi, or Assoc. Prof. DI Dr. Klaus Schoeffmann.

Besides, a reference must be made to the following publication [1] when this dataset is used in any academic and research reports:

Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., El-Shabrawi, Y.,Schoeffmann, K.: LensID: A CNN-RNN-Based Framework Towards Lens Irregularity Detection in Cataract Surgery Videos. In: 24th International Conference on Medical Image Computing & Computer Assisted Inter-ventions (MICCAI 2021). (to appear)

The datasets are licensed under Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0, Creative Commons License) and are created as well as maintained by Distributed Multimedia Systems Group of the Institute of Information Technology (ITEC) at Alpen-Adria Universit├Ąt in Klagenfurt, Austria.

This license allows users of this dataset to copy, distribute, and transmit the work under the following conditions:

For further legal details, please read the complete license terms.


Download

If you agree to above conditions, you are free to download:


References

[1] Ghamsarian, N., Taschwer, M., Putzgruber-Adamitsch, D., Sarny, S., El-Shabrawi, Y.,Schoeffmann, K.: LensID: A CNN-RNN-Based Framework Towards Lens Irregularity Detection in Cataract Surgery Videos. In: 24th International Conference on Medical Image Computing & Computer Assisted Inter-ventions (MICCAI 2021). (to appear)