University of Twente Student Theses


Autonomous anatomical structure recognition using machine learning

Schuhmacher, M. (2018) Autonomous anatomical structure recognition using machine learning.

[img] PDF
Abstract:Introduction. Object detection algorithms are already applied to tomographical medical data (i.e. X-Ray scans), but not to video data of intra-abdominal minimally invasive surgery. This research made a start to work on anatomical structure recognition using machine learning algorithms. Method. The YOLOv2 algorithm was applied to a dataset of 8 videos of colorectal surgery, with the goal to identify five anatomical structures (ureter, tendon, artery, white line of Toldt, colon) purely on visual information. 7 videos (6189 images) were used for training and validation, 1 video (1185 images) was used a test set, all annotated with bounding boxes around the target structures. Training parameters with the Adam optimizer were α = 0.0001 and batch size 8, trained for 100 epochs with a checkpoint on the lowest validation loss. A 3-layer LSTM network was added after the YOLOv2 algorithm for better performance, which was trained at α = 0.00001 for 300 epochs with a batch size of 32 and dropout of 0.2. Results. The LSTM implementation failed to produce reliable results, with indications that the loss function was implemented falsely. The standalone YOLOv2 network had a mean average precision (mAP) of 43.72% on our test set, which is comparable to its performance on the standard object detection dataset COCO. Discussion. The faulty LSTM implementation was narrowed down to probably a fault in the loss function. Annotation in medical data requires skill, anatomical background, and time, and is still difficult and inconsistent. Performance of YOLOv2 can be improved by training on a larger dataset, removing the ‘Toldt’ class, correctly implement the LSTM network, and keep testing for optimal settings in thresholds and learning parameters. Conclusion. Object detection algorithms can be applied to video data of minimally invasive abdominal surgery with an acceptable precision.
Item Type:Essay (Master)
Faculty:TNW: Science and Technology
Subject:44 medicine, 50 technical science in general, 54 computer science
Programme:Technical Medicine MSc (60033)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page