Abstract
Background
Our previous work classified a taxonomy of needle driving gestures during a vesicourethral anastomosis of robotic radical prostatectomy in association with tissue tears and patient outcomes. Herein we train deep-learning based computer vision (CV) to automate the identification and classification of suturing gestures for needle driving attempts.
Methods
Two independent raters manually annotated live suturing video clips to label time points and gestures. Identification (2,395 videos) and classification (511 videos) datasets were compiled to train CV models to produce two- and five-class label predictions, respectively. Networks were trained on inputs of raw RGB pixels as well as optical flow for each frame. We explore the effect of different recurrent models (LSTM vs. convLSTM). All models were trained on 80/20 train/test splits.
Results
We observe that all models are able to reliably predict either the presence of a gesture (identification, AUC: 0.88) as well as the type of gesture (classification, AUC: 0.87) at significantly above chance levels. For both gesture identification and classification datasets, we observed no effect of recurrent classification model choice on performance.
Conclusions
Our results demonstrate CV’s ability to recognize features that not only can identify the action of suturing but also distinguish between different classifications of suturing gestures. This demonstrates the potential to utilize deep learning CV towards future automation of surgical skill assessment.
Article Summary:
This study utilizes deep learning-based computer vision to identify specific moments of surgical suturing activity and to classify each specific suturing gesture applied during robot-assisted surgery. The importance of this work is the foundation it provides for future automation of surgical skill assessment for training feedback.
Introduction
Growing evidence supports that superior surgical performance is associated with superior clinical outcomes.1,2 Yet how we presently assess surgery --manual evaluation by peers -- is fraught with subjectivity and is not scalable.3,4
Tremendous work has been done already to better assess surgeon performance during robot-assisted surgeries. For example, with suturing, the robotic anastomosis competency evaluation (RACE) has been developed to streamline technical skills assessment with objective criteria for each suturing skill domain5. Yet even with such a rubric, manual assessment and feedback of every suture performed by a training surgeon is not feasible. Our group previously deconstructed suturing into a clinically meaningful manner to consist of 3 phases (needle position, needle driving, and suture cinching; Fig 1), and further developed a classification system for suturing gestures to standardize the training and assessment of robot-assisted suturing (Fig 2)6. We have demonstrated that surgeon selection of gestures at specific anatomic positions during the vesico-urethral anastomosis (VUA) during the robot-assisted radical prostatectomy (RARP) is linked to surgeon efficiency and clinical outcomes (i.e., tissue tear)6. We have also demonstrated that when surgeons are instructed on what specific gesture to utilize during the VUA, they are able to shorten the learning curve for this step of the RARP7.
Computational approaches have already been tapped towards the goal of recognizing and evaluating surgical gestures. Classical computer vision techniques8, as well as recurrent models using kinematics9 have been employed previously with modest success. In recent years, neural networks for extracting information from video data have made tremendous strides.10,11 Indeed, some groups have started to apply such deep learning approaches to commonly available datasets such as the JIGSAWS suture classification dataset.12 While these prior works have been largely limited to the well-controlled laboratory environment, live application of computer vision-based identification and classification of suturing gestures will ultimately determine the real-world utility of such technology.
Herein, we utilize deep learning-based computer vision to 1) identify suture needle driving activity during live robot-assisted surgery; 2) classify suturing needle driving gestures based on a clinically validated categorization we previously described.
Methods
In this study, we set out to characterize commonly used architectures employed in action recognition towards the goal of recognizing and classifying surgical stitches. To undertake this study, we started by generating two complementary datasets for training models from videos of a live VUA during a RARP to identify when a suturing gesture is happening (gesture identification) and what gesture is happening (gesture classification). Using annotated video data from a previous study,6 we generated a dataset of short clips corresponding to moments of “needle driving” (Fig 1b) (positive samples) and short clips corresponding to non-needle driving surgical activity (negative samples). This dataset which we call the “identification dataset” contained 2,395 total video clips (1209 positive; 1186 negative) with an average duration of 12.2 seconds. For gesture classification, we generated a dataset of 511 total clips to distinguish five selected gestures from our established taxonomy (Fig 2). These five were selected based on the adequate sample size per class (Gesture 1 – 150 samples, 2–101, 3–96, 4–117, 5–47). The clips had an average duration of 6.6 seconds and each one was manually labeled by two independent trained annotators. We refer to this dataset as the “classification dataset”.
The computational task of identifying actions from video inputs is commonly known in computer vision as action recognition. Although a challenging problem, neural networks have recently shown promise in their ability to reason from such spatiotemporal data. The most common example of such networks is so-called “two-stream networks” in which networks take two streams of inputs as features: the raw RGB pixels of the video as well as an optical flow representation in which momentary direction and magnitude of motion are defined at each pixel (Fig 3). These inputs are usually passed through a standard feature extractor (usually a deep network) and the representations produced by these networks are further passed into a temporally recurrent classification layer, usually some flavor of a long short term memory unit (LSTM13). In practice, one can add complexity or inductive biases to the recurrent classification for example by making this layer convolutional (convLSTM14), which may aid in performance and training time. In this work, we explore specific hyper parameter choices in this framework for the recurrent classification model (Figure 2). For the comparisons presented here we chose a fixed 7-layer network (AlexNet15), which was initialized from weights trained on a large corpus of natural images (ImageNet). We vary the recurrent classification layer (LSTM, convLSTM) in our experiments.
Using the two curated datasets as our starting point, we set out to evaluate commonly used deep learning architectures used in action recognition for the task of identifying when (identification) and what (classification) suturing gestures happened. Taken together, we hope this work serves as a preliminary demonstration of a potential approach towards merging the latest research in deep learning with the identification, classification, and potential evaluation of surgical skills to improve patient outcomes.
Results
We started by training a model to identify short clips as either containing “needle driving” (positive label) or did not contain such an action (negative label) using the annotated identification dataset. We trained all models on three 80/20 train/test splits, using hyper parameters shown in Table 1 and report AUC and accuracy in Figure 4. We observed significantly above chance values for both accuracy (79%) and AUC (0.88) in the identification task, however we found no effect of recurrent classification model on the model performance.
Table 1:
Hyperparameter | Value (classification) | Value (cutting) |
---|---|---|
N classes | 2 | 5 |
Learning algorithm | Adam | Adam |
Learning rate | 1e-5 | 1e-5 |
Epochs | 25 | 7 |
Batch size | 1 | 1 |
Base network | {AlexNet} | {AlexNet} |
Classification network | {LSTM, convLSTM} | {LSTM, convLSTM} |
For each network, 3 different train/test split networks were trained. For video clips longer than 4 sec, a random 4-sec clip was grabbed on each iteration. Classes were balanced via upsampling during training. For data augmentation each frame was resized (240×240) and cropped (224×224) during training. A stride of 4 frames was used (e.g. only sample 1 out of every 4 frames) for an effective frame rate of 7.5Hz. Images were standard scaled before passing through the network. LSTM was a 2-layer LSTM with 64 and 128 units in each layer, respectively. Convolutional LSTM had stride 3 and the number of channels equal to the number of channels in the final convolutional layer of the respective base network (256 for AlexNet).
We further moved on to train a model for identifying when a gesture happened using the classification dataset to output gesture type probabilities over the 5 selected gestures in our dataset (Figure 2). We varied the same hyper parameters as before (classification layer) and found that similar to the identification task that there was no effect of the specific type of classification model. We do however note that convolutional versions of the LSTM (convLSTM) reached convergence in fewer epochs than LSTM counterparts (data not shown). In this classification task, we achieved an average 1st guess (top1) accuracy of 62% for the models trained. Additionally, we also managed to maintain a high AUC (0.87), indicating that the model does not take a biased approach to the classification task to achieve good results. This is further evident in the confusion matrix in Figure 5, where a strong diagonal is present, indicative of reasonable performance in all classes.
Discussion
In summary, we present a novel annotated dataset for the study of suture gestures in the context of a robot-assisted surgical procedure. We produced annotation for two types of tasks, one with clips annotated with when “needle driving” is present (gesture identification dataset) and another dataset labeled with gesture clips and their corresponding type according to the presented taxonomy (gesture classification dataset). We further show that applying standard deep network approaches, commonly used in action recognition, can be used to train models that achieve promising performance on both tasks.
The results presented here, in many ways, present a conservative estimate of the sort of performance that can be achieved from these models. We are training in a relatively data-limited regime in both tasks so these models will further improve as labeled data becomes available. In addition, we did not yet employ any inference “tricks” such as ensemble or majority votes commonly used in action recognition models.10,16
Our present study is foundational to future work on automating technical skills evaluation. Having completed the first steps to identify and classify suturing gestures, we will transition to evaluating how well a suture is executed. Part of how well suturing is performed is simply gesture selection at specific anatomic positions6, in which the present study can help streamline. But the suturing performance also depends on the actual technical skill of the surgeon in carrying out the maneuver, and the models we develop in this study hold promise for such automatic evaluation as well.
On a higher level, our present work is foundational not only for evaluation of suturing, but it also builds the starting point for eventual autonomous suturing. Such future platforms must first be capable of recognizing and assessing ideal suturing skills before becoming capable of performing it autonomously.
Table 2:
Number of trainable parameters | |||
---|---|---|---|
Feature extractor | Recurrent model | Total | |
AlexNet-LSTM | 3.7 million | 49k | 7.5 million |
AlexNet-convLSTM | 3.7 million | 7 million | 14.4 million |
Total number of trainable parameters for each of the 2 configurations trained. Note that more parameters does not necessarily improve performance.
Funding/Support:
This study is supported in part by the National Institute Of Biomedical Imaging And Bioengineering of the National Institutes of Health under Award Number K23EB026493.
Footnotes
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
COI/Disclosure: Andrew J. Hung has financial disclyosures with Quantgene, Inc. (consultant), Mimic Technologies, Inc. (consultant), and Johnson & Johnson (consultant).
References
- 1.Birkmeyer JD, Finks JF, O’Reilly A, Oerline M, Carlin AM, Nunn AR, et al. Surgical skill and complication rates after bariatric surgery. N Engl J Med. 2013;369(15):1434–42. [DOI] [PubMed] [Google Scholar]
- 2.Hung AJ, Chen J, Ghodoussipour S, Oh PJ, Liu Z, Nguyen J, et al. A deep-learning model using automated performance metrics and clinical features to predict urinary continence recovery after robot-assisted radical prostatectomy. BJU Int. 2019;124(3):487–495. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Aghazadeh MA, Jayaratna IS, Hung AJ, Pan MM, Desai MM, Gill IS, et al. External validation of Global Evaluative Assessment of Robotic Skills (GEARS). Surg Endosc. 2015;29(11):3261–6. [DOI] [PubMed] [Google Scholar]
- 4.Hung AJ, Chen J, Jarc A, Hatcher D, Djaladat H, Gill IS. Development and Validation of Objective Performance Metrics for Robot-Assisted Radical Prostatectomy: A Pilot Study. J Urol. 2018;199(1):296–304. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Raza SJ, Field E, Jay C, Eun D, Fumo M, Hu JC, et al. Surgical competency for urethrovesical anastomosis during robot-assisted radical prostatectomy: development and validation of the robotic anastomosis competency evaluation. Urology. 2015;85(1):27–32. [DOI] [PubMed] [Google Scholar]
- 6.Chen J, Oh PJ, Cheng N, Shah A, Montez J, Jarc A, et al. Use of Automated Performance Metrics to Measure Surgeon Performance during Robotic Vesicourethral Anastomosis and Methodical Development of a Training Tutorial. J Urol. 2018;200(4):895–902. [DOI] [PubMed] [Google Scholar]
- 7.Remulla D, Nguyen J, Lee R, Chu T, Chen J, Singh V, Chen A, Marshall S, Gill I, Hung AJ (2019) Impact of a standardized training tutorial on automated performance metrics and cognitive workload during robotic vesicourethral anastomosis. J Urol. 2019;201(4): e509. [Google Scholar]
- 8.Zappella L, Bejar B, Hager G, Vidal R. Surgical gesture classification from video and kinematic data. Med Image Anal. 2013;17(7):732–45. [DOI] [PubMed] [Google Scholar]
- 9.DiPietro R, Ahmidi N, Malpani A, Waldram M, Lee GI, Lee MR, et al. Segmenting and classifying activities in robot-assisted surgery with recurrent neural networks. Int J Comput Assist Radiol Surg. 2019;14(11):2005–20. [DOI] [PubMed] [Google Scholar]
- 10.Feichtenhofer C, Fan H, Malik J, He K. SlowFast Networks for Video Recognition. 2018;arXiv:1812.03982. [Google Scholar]
- 11.Feichtenhofer C, Pinz A, Zisserman A. Convolutional Two-Stream Network Fusion for Video Action Recognition. 2016;arXiv:1604.06573. [Google Scholar]
- 12.[dataset] Gao Y, Vedula S, Reiley C, Ahmidi N, Varadarajan B, Lin H, et al. The JHU-ISI Gesture and Skill Assessment Working Set (JIGSAWS): A Surgical Activity Dataset for Human Motion Modeling, In Modeling and Monitoring of Computer Assisted Interventions (M2CAI) – MICCAI Workshop, 2014. [Google Scholar]
- 13.Hochreiter S, Schmidhuber J. Long Short-Term Memory. Neural Computation. 1997; 9(8): 1735–80. [DOI] [PubMed] [Google Scholar]
- 14.Shi X, Chen Z, Wang H, Yeung DY, Wong W, Woo W. Convolutional LSTM Network: A Machine Learning Approach for Precipitation Nowcasting. 2015; arXiv:1506.04214. [Google Scholar]
- 15.Krizhevsky A, Sutskever I, Hinton G. ImageNet classification with deep convolutional neural networks. Proceedings of the 25th International Conference on Neural Information Processing Systems. 2012; 1:1097–1105. [Google Scholar]
- 16.Tran D, Bourdev L, Fergus R, Torresani L, Paluri M. Learning Spatiotemporal Features with 3D Convolutional Networks. 2015;arXiv:1412.0767 [Google Scholar]