Expert surgeons and deep learning models can predict the outcome of surgical hemorrhage from 1 min of video.

Major vascular injury resulting in uncontrolled bleeding is a catastrophic and often fatal complication of minimally invasive surgery. At the outset of these events, surgeons do not know how much blood will be lost or whether they will successfully control the hemorrhage (achieve hemostasis). We evaluate the ability of a deep learning neural network (DNN) to predict hemostasis control ability using the first minute of surgical video and compare model performance with human experts viewing the same video. The publicly available SOCAL dataset contains 147 videos of attending and resident surgeons managing hemorrhage in a validated, high-fidelity cadaveric simulator. Videos are labeled with outcome and blood loss (mL). The first minute of 20 videos was shown to four, blinded, fellowship trained skull-base neurosurgery instructors, and to SOCALNet (a DNN trained on SOCAL videos). SOCALNet architecture included a convolutional network (ResNet) identifying spatial features and a recurrent network identifying temporal features (LSTM). Experts independently assessed surgeon skill, predicted outcome and blood loss (mL). Outcome and blood loss predictions were compared with SOCALNet. Expert inter-rater reliability was 0.95. Experts correctly predicted 14/20 trials (Sensitivity: 82%, Specificity: 55%, Positive Predictive Value (PPV): 69%, Negative Predictive Value (NPV): 71%). SOCALNet correctly predicted 17/20 trials (Sensitivity 100%, Specificity 66%, PPV 79%, NPV 100%) and correctly identified all successful attempts. Expert predictions of the highest and lowest skill surgeons and expert predictions reported with maximum confidence were more accurate. Experts systematically underestimated blood loss (mean error - 131 mL, RMSE 350 mL, R2 0.70) and fewer than half of expert predictions identified blood loss > 500 mL (47.5%, 19/40). SOCALNet had superior performance (mean error - 57 mL, RMSE 295 mL, R2 0.74) and detected most episodes of blood loss > 500 mL (80%, 8/10). In validation experiments, SOCALNet evaluation of a critical on-screen surgical maneuver and high/low-skill composite videos were concordant with expert evaluation. Using only the first minute of video, experts and SOCALNet can predict outcome and blood loss during surgical hemorrhage. Experts systematically underestimated blood loss, and SOCALNet had no false negatives. DNNs can provide accurate, meaningful assessments of surgical video. We call for the creation of datasets of surgical adverse events for quality improvement research.

Utility of the Simulated Outcomes Following Carotid Artery Laceration Video Data Set for Machine Learning Applications.

Importance: Surgical data scientists lack video data sets that depict adverse events, which may affect model generalizability and introduce bias. Hemorrhage may be particularly challenging for computer vision-based models because blood obscures the scene. Objective: To assess the utility of the Simulated Outcomes Following Carotid Artery Laceration (SOCAL)-a publicly available surgical video data set of hemorrhage complication management with instrument annotations and task outcomes-to provide benchmarks for surgical data science techniques, including computer vision instrument detection, instrument use metrics and outcome associations, and validation of a SOCAL-trained neural network using real operative video. Design, Setting, and Participants: For this quailty improvement study, a total of 75 surgeons with 1 to 30 years' experience (mean, 7 years) were filmed from January 1, 2017, to December 31, 2020, managing catastrophic surgical hemorrhage in a high-fidelity cadaveric training exercise at nationwide training courses. Videos were annotated from January 1 to June 30, 2021. Interventions: Surgeons received expert coaching between 2 trials. Main Outcomes and Measures: Hemostasis within 5 minutes (task success, dichotomous), time to hemostasis (in seconds), and blood loss (in milliliters) were recorded. Deep neural networks (DNNs) were trained to detect surgical instruments in view. Model performance was measured using mean average precision (mAP), sensitivity, and positive predictive value. Results: SOCAL contains 31 443 frames with 65 071 surgical instrument annotations from 147 trials with associated surgeon demographic characteristics, time to hemostasis, and recorded blood loss for each trial. Computer vision-based instrument detection methods using DNNs trained on SOCAL achieved a mAP of 0.67 overall and 0.91 for the most common surgical instrument (suction). Hemorrhage control challenges standard object detectors: detection of some surgical instruments remained poor (mAP, 0.25). On real intraoperative video, the model achieved a sensitivity of 0.77 and a positive predictive value of 0.96. Instrument use metrics derived from the SOCAL video were significantly associated with performance (blood loss). Conclusions and Relevance: Hemorrhage control is a high-stakes adverse event that poses unique challenges for video analysis, but no data sets of hemorrhage control exist. The use of SOCAL, the first data set to depict hemorrhage control, allows the benchmarking of data science applications, including object detection, performance metric development, and identification of metrics associated with outcomes. In the future, SOCAL may be used to build and validate surgical data science models.

Prevalence of neural collapse during the terminal phase of deep learning training.

Modern practice for training classification deepnets involves a terminal phase of training (TPT), which begins at the epoch where training error first vanishes. During TPT, the training error stays effectively zero, while training loss is pushed toward zero. Direct measurements of TPT, for three prototypical deepnet architectures and across seven canonical classification datasets, expose a pervasive inductive bias we call neural collapse (NC), involving four deeply interconnected phenomena. (NC1) Cross-example within-class variability of last-layer training activations collapses to zero, as the individual activations themselves collapse to their class means. (NC2) The class means collapse to the vertices of a simplex equiangular tight frame (ETF). (NC3) Up to rescaling, the last-layer classifiers collapse to the class means or in other words, to the simplex ETF (i.e., to a self-dual configuration). (NC4) For a given activation, the classifier's decision collapses to simply choosing whichever class has the closest train class mean (i.e., the nearest class center [NCC] decision rule). The symmetric and very simple geometry induced by the TPT confers important benefits, including better generalization performance, better robustness, and better interpretability.

Multi-Scale Patch-Based Image Restoration.

Many image restoration algorithms in recent years are based on patch processing. The core idea is to decompose the target image into fully overlapping patches, restore each of them separately, and then merge the results by a plain averaging. This concept has been demonstrated to be highly effective, leading often times to the state-of-the-art results in denoising, inpainting, deblurring, segmentation, and other applications. While the above is indeed effective, this approach has one major flaw: the prior is imposed on intermediate (patch) results, rather than on the final outcome, and this is typically manifested by visual artifacts. The expected patch log likelihood (EPLL) method by Zoran and Weiss was conceived for addressing this very problem. Their algorithm imposes the prior on the patches of the final image, which in turn leads to an iterative restoration of diminishing effect. In this paper, we propose to further extend and improve the EPLL by considering a multi-scale prior. Our algorithm imposes the very same prior on different scale patches extracted from the target image. While all the treated patches are of the same size, their footprint in the destination image varies due to subsampling. Our scheme comes to alleviate another shortcoming existing in patch-based restoration algorithms--the fact that a local (patch-based) prior is serving as a model for a global stochastic phenomenon. We motivate the use of the multi-scale EPLL by restricting ourselves to the simple Gaussian case, comparing the aforementioned algorithms and showing a clear advantage to the proposed method. We then demonstrate our algorithm in the context of image denoising, deblurring, and super-resolution, showing an improvement in performance both visually and quantitatively.