Luis C. Garcia-Peraza-Herrera

Charlie Budd, Luis C. Garcia-Peraza-Herrera, Martin Huber et al.

Endoscopic content area refers to the informative area enclosed by the dark, non-informative, border regions present in most endoscopic footage. The estimation of the content area is a common task in endoscopic image processing and computer vision pipelines. Despite the apparent simplicity of the problem, several factors make reliable real-time estimation surprisingly challenging. The lack of rigorous investigation into the topic combined with the lack of a common benchmark dataset for this task has been a long-lasting issue in the field. In this paper, we propose two variants of a lean GPU-based computational pipeline combining edge detection and circle fitting. The two variants differ by relying on handcrafted features, and learned features respectively to extract content area edge point candidates. We also present a first-of-its-kind dataset of manually annotated and pseudo-labelled content areas across a range of surgical indications. To encourage further developments, the curated dataset, and an implementation of both algorithms, has been made public (https://doi.org/10.7303/syn32148000, https://github.com/charliebudd/torch-content-area). We compare our proposed algorithm with a state-of-the-art U-Net-based approach and demonstrate significant improvement in terms of both accuracy (Hausdorff distance: 6.3 px versus 118.1 px) and computational time (Average runtime per frame: 0.13 ms versus 11.2 ms).

Luis C. Garcia-Peraza-Herrera, Sebastien Ourselin, Tom Vercauteren

The performance of deep learning (DL) algorithms is heavily influenced by the quantity and the quality of the annotated data. However, in Surgical Data Science, access to it is limited. It is thus unsurprising that substantial research efforts are made to develop methods aiming at mitigating the scarcity of annotated SDS data. In parallel, an increasing number of Computer Assisted Interventions (CAI) datasets are being released, although the scale of these remain limited. On these premises, data curation is becoming a key element of many SDS research endeavors. Surgical video datasets are demanding to curate and would benefit from dedicated support tools. In this work, we propose to summarize surgical videos into storyboards or collages of representative frames to ease visualization, annotation, and processing. Video summarization is well-established for natural images. However, state-of-the-art methods typically rely on models trained on human-made annotations, few methods have been evaluated on surgical videos, and the availability of software packages for the task is limited. We present videosum, an easy-to-use and open-source Python library to generate storyboards from surgical videos that contains a variety of unsupervised methods.

Luis C. Garcia-Peraza-Herrera, Conor Horgan, Sebastien Ourselin et al.

Visual discrimination of clinical tissue types remains challenging, with traditional RGB imaging providing limited contrast for such tasks. Hyperspectral imaging (HSI) is a promising technology providing rich spectral information that can extend far beyond three-channel RGB imaging. Moreover, recently developed snapshot HSI cameras enable real-time imaging with significant potential for clinical applications. Despite this, the investigation into the relative performance of HSI over RGB imaging for semantic segmentation purposes has been limited, particularly in the context of medical imaging. Here we compare the performance of state-of-the-art deep learning image segmentation methods when trained on hyperspectral images, RGB images, hyperspectral pixels (minus spatial context), and RGB pixels (disregarding spatial context). To achieve this, we employ the recently released Oral and Dental Spectral Image Database (ODSI-DB), which consists of 215 manually segmented dental reflectance spectral images with 35 different classes across 30 human subjects. The recent development of snapshot HSI cameras has made real-time clinical HSI a distinct possibility, though successful application requires a comprehensive understanding of the additional information HSI offers. Our work highlights the relative importance of spectral resolution, spectral range, and spatial information to both guide the development of HSI cameras and inform future clinical HSI applications.

Chengan Che, Chao Wang, Jiayuan Huang et al.

Recent advancements in self-supervised learning have led to powerful surgical vision encoders capable of spatiotemporal understanding. However, extending these visual foundations to multi-modal reasoning tasks is severely bottlenecked by the prohibitive cost of expert textual annotations. To overcome this scalability limitation, we introduce \textbf{LIME}, a large-scale multi-modal dataset derived from open-access surgical videos using human-free, Large Language Model (LLM)-generated narratives. While LIME offers immense scalability, unverified generated texts may contain errors, including hallucinations, that could potentially lead to catastrophically degraded pre-trained medical priors in standard contrastive pipelines. To mitigate this, we propose \textbf{SurgLIME}, a parameter-efficient Vision-Language Pre-training (VLP) framework designed to learn reliable cross-modal alignments using noisy narratives. SurgLIME preserves foundational medical priors using a LoRA-adapted dual-encoder architecture and introduces an automated confidence estimation mechanism that dynamically down-weights uncertain text during contrastive alignment. Evaluations on the AutoLaparo and Cholec80 benchmarks show that SurgLIME achieves competitive zero-shot cross-modal alignment while preserving the robust linear probing performance of the visual foundation model. Dataset, code, and models are publicly available at \href{https://github.com/visurg-ai/SurgLIME}{https://github.com/visurg-ai/SurgLIME}.

Meng Wei, Charlie Budd, Luis C. Garcia-Peraza-Herrera et al.

Surgical instrument segmentation is recognised as a key enabler in providing advanced surgical assistance and improving computer-assisted interventions. In this work, we propose SegMatch, a semi-supervised learning method to reduce the need for expensive annotation for laparoscopic and robotic surgical images. SegMatch builds on FixMatch, a widespread semi supervised classification pipeline combining consistency regularization and pseudo-labelling, and adapts it for the purpose of segmentation. In our proposed SegMatch, the unlabelled images are first weakly augmented and fed to the segmentation model to generate pseudo-labels. In parallel, images are fed to a strong augmentation branch and consistency between the branches is used as an unsupervised loss. To increase the relevance of our strong augmentations, we depart from using only handcrafted augmentations and introduce a trainable adversarial augmentation strategy. Our FixMatch adaptation for segmentation tasks further includes carefully considering the equivariance and invariance properties of the augmentation functions we rely on. For binary segmentation tasks, our algorithm was evaluated on the MICCAI Instrument Segmentation Challenge datasets, Robust-MIS 2019 and EndoVis 2017. For multi-class segmentation tasks, we relied on the recent CholecInstanceSeg dataset. Our results show that SegMatch outperforms fully-supervised approaches by incorporating unlabelled data, and surpasses a range of state-of-the-art semi-supervised models across different labelled to unlabelled data ratios.

Theodore Barfoot, Luis C. Garcia-Peraza-Herrera, Samet Akcay et al.

Deep neural networks for medical image segmentation are often overconfident, compromising both reliability and clinical utility. In this work, we propose differentiable formulations of marginal L1 Average Calibration Error (mL1-ACE) as an auxiliary loss that can be computed on a per-image basis. We compare both hard- and soft-binning approaches to directly improve pixel-wise calibration. Our experiments on four datasets (ACDC, AMOS, KiTS, BraTS) demonstrate that incorporating mL1-ACE significantly reduces calibration errors, particularly Average Calibration Error (ACE) and Maximum Calibration Error (MCE), while largely maintaining high Dice Similarity Coefficients (DSCs). We find that the soft-binned variant yields the greatest improvements in calibration, over the Dice plus cross-entropy loss baseline, but often compromises segmentation performance, with hard-binned mL1-ACE maintaining segmentation performance, albeit with weaker calibration improvement. To gain further insight into calibration performance and its variability across an imaging dataset, we introduce dataset reliability histograms, an aggregation of per-image reliability diagrams. The resulting analysis highlights improved alignment between predicted confidences and true accuracies. Overall, our approach not only enhances the trustworthiness of segmentation predictions but also shows potential for safer integration of deep learning methods into clinical workflows. We share our code here: https://github.com/cai4cai/Average-Calibration-Losses

Chengan Che, Chao Wang, Tom Vercauteren et al.

Traditional open-access datasets focusing on surgical procedures are often limited by their small size, typically consisting of fewer than 100 videos and less than 30 hours of footage, which leads to poor model generalization. To address this constraint, a new dataset called LEMON has been compiled using a novel aggregation pipeline that collects high-resolution videos from online sources. Featuring an extensive collection of over 4K surgical videos totaling 938 hours (85 million frames) of high-quality footage across multiple procedure types, LEMON offers a comprehensive resource surpassing existing alternatives in size and scope, including two novel downstream tasks. To demonstrate the effectiveness of this diverse dataset, we introduce LemonFM, a foundation model pretrained on LEMON using a novel self-supervised augmented knowledge distillation approach. LemonFM consistently outperforms existing surgical foundation models across four downstream tasks and six datasets, achieving significant gains in surgical phase recognition (+9.5pp, +9.4pp, and +8.4pp of Jaccard in AutoLaparo, M2CAI16, and Cholec80), surgical action recognition (+4.4pp of mAP in CholecT50), surgical tool presence detection (+5.3pp and +10.2pp of mAP in Cholec80 and GraSP), and surgical semantic segmentation (+8.3pp of mDice in CholecSeg8k). LEMON and LemonFM will serve as foundational resources for the research community and industry, accelerating progress in developing autonomous robotic surgery systems and ultimately contributing to safer and more accessible surgical care worldwide.

Chao Wang, Chengan Che, Xinyue Chen et al.

Counterfactual explanations (CFEs) are minimal and semantically meaningful modifications of the input of a model that alter the model predictions. They highlight the decisive features the model relies on, providing contrastive interpretations for classifiers. State-of-the-art visual counterfactual explanation methods are designed to explain image classifiers. The generation of CFEs for video classifiers remains largely underexplored. For the counterfactual videos to be useful, they have to be physically plausible, temporally coherent, and exhibit smooth motion trajectories. Existing CFE image-based methods, designed to explain image classifiers, lack the capacity to generate temporally coherent, smooth and physically plausible video CFEs. To address this, we propose Back To The Feature (BTTF), an optimization framework that generates video CFEs. Our method introduces two novel features, 1) an optimization scheme to retrieve the initial latent noise conditioned by the first frame of the input video, 2) a two-stage optimization strategy to enable the search for counterfactual videos in the vicinity of the input video. Both optimization processes are guided solely by the target classifier, ensuring the explanation is faithful. To accelerate convergence, we also introduce a progressive optimization strategy that incrementally increases the number of denoising steps. Extensive experiments on video datasets such as Shape-Moving (motion classification), MEAD (emotion classification), and NTU RGB+D (action classification) show that our BTTF effectively generates valid, visually similar and realistic counterfactual videos that provide concrete insights into the classifier's decision-making mechanism.

Chengan Che, Chao Wang, Xinyue Chen et al.

Procedural activities, ranging from routine cooking to complex surgical operations, are highly structured as a set of actions conducted in a specific temporal order. Despite their success on static images and short clips, current self-supervised learning methods often overlook the procedural nature that underpins such activities. We expose the lack of procedural awareness in current SSL methods with a motivating experiment: models pretrained on forward and time-reversed sequences produce highly similar features, confirming that their representations are blind to the underlying procedural order. To address this shortcoming, we propose PL-Stitch, a self-supervised framework that harnesses the inherent temporal order of video frames as a powerful supervisory signal. Our approach integrates two novel probabilistic objectives based on the Plackett-Luce (PL) model. The primary PL objective trains the model to sort sampled frames chronologically, compelling it to learn the global workflow progression. The secondary objective, a spatio-temporal jigsaw loss, complements the learning by capturing fine-grained, cross-frame object correlations. Our approach consistently achieves superior performance across five surgical and cooking benchmarks. Specifically, PL-Stitch yields significant gains in surgical phase recognition (e.g., +11.4 pp k-NN accuracy on Cholec80) and cooking action segmentation (e.g., +5.7 pp linear probing accuracy on Breakfast), demonstrating its effectiveness for procedural video representation learning.

Nikolay Manchev, Luis C. Garcia-Peraza-Herrera

Recurrent Neural Networks (RNNs) are commonly used for real-time processing, streaming data, and cases where the amount of training samples is limited. Backpropagation Through Time (BPTT) is the predominant algorithm for training RNNs; however, it is frequently criticized for being prone to exploding and vanishing gradients and being biologically implausible. In this paper, we present and evaluate a target propagation-based method for RNNs, which uses local updates and seeks to reduce the said instabilities. Having stable RNN models increases their practical use in a wide range of fields such as natural language processing, time-series forecasting, anomaly detection, control systems, and robotics. The proposed solution uses local representation alignment (LRA). We thoroughly analyze the performance of this method, experiment with normalization and different local error functions, and invalidate certain assumptions about the behavior of this type of learning. Namely, we demonstrate that despite the decomposition of the network into sub-graphs, the model still suffers from vanishing gradients. We also show that gradient clipping as proposed in LRA has little to no effect on network performance. This results in an LRA RNN model that is very difficult to train due to vanishing gradients. We address this by introducing gradient regularization in the direction of the update and demonstrate that this modification promotes gradient flow and meaningfully impacts convergence. We compare and discuss the performance of the algorithm, and we show that the regularized LRA RNN considerably outperforms the unregularized version on three landmark tasks: temporal order, 3-bit temporal order, and random permutation.

Yang Liu, Maxence Boels, Luis C. Garcia-Peraza-Herrera et al.

Online surgical phase recognition plays a significant role towards building contextual tools that could quantify performance and oversee the execution of surgical workflows. Current approaches are limited since they train spatial feature extractors using frame-level supervision that could lead to incorrect predictions due to similar frames appearing at different phases, and poorly fuse local and global features due to computational constraints which can affect the analysis of long videos commonly encountered in surgical interventions. In this paper, we present a two-stage method, called Long Video Transformer (LoViT) for fusing short- and long-term temporal information that combines a temporally-rich spatial feature extractor and a multi-scale temporal aggregator consisting of two cascaded L-Trans modules based on self-attention, followed by a G-Informer module based on ProbSparse self-attention for processing global temporal information. The multi-scale temporal head then combines local and global features and classifies surgical phases using phase transition-aware supervision. Our approach outperforms state-of-the-art methods on the Cholec80 and AutoLaparo datasets consistently. Compared to Trans-SVNet, LoViT achieves a 2.4 pp (percentage point) improvement in video-level accuracy on Cholec80 and a 3.1 pp improvement on AutoLaparo. Moreover, it achieves a 5.3 pp improvement in phase-level Jaccard on AutoLaparo and a 1.55 pp improvement on Cholec80. Our results demonstrate the effectiveness of our approach in achieving state-of-the-art performance of surgical phase recognition on two datasets of different surgical procedures and temporal sequencing characteristics whilst introducing mechanisms that cope with long videos.

Caspar Gruijthuijsen, Luis C. Garcia-Peraza-Herrera, Gianni Borghesan et al.

Many keyhole interventions rely on bi-manual handling of surgical instruments, forcing the main surgeon to rely on a second surgeon to act as a camera assistant. In addition to the burden of excessively involving surgical staff, this may lead to reduced image stability, increased task completion time and sometimes errors due to the monotony of the task. Robotic endoscope holders, controlled by a set of basic instructions, have been proposed as an alternative, but their unnatural handling may increase the cognitive load of the (solo) surgeon, which hinders their clinical acceptance. More seamless integration in the surgical workflow would be achieved if robotic endoscope holders collaborated with the operating surgeon via semantically rich instructions that closely resemble instructions that would otherwise be issued to a human camera assistant, such as "focus on my right-hand instrument". As a proof of concept, this paper presents a novel system that paves the way towards a synergistic interaction between surgeons and robotic endoscope holders. The proposed platform allows the surgeon to perform a bimanual coordination and navigation task, while a robotic arm autonomously performs the endoscope positioning tasks. Within our system, we propose a novel tooltip localization method based on surgical tool segmentation and a novel visual servoing approach that ensures smooth and appropriate motion of the endoscope camera. We validate our vision pipeline and run a user study of this system. The clinical relevance of the study is ensured through the use of a laparoscopic exercise validated by the European Academy of Gynaecological Surgery which involves bi-manual coordination and navigation. Successful application of our proposed system provides a promising starting point towards broader clinical adoption of robotic endoscope holders.

Luis C. Garcia-Peraza-Herrera, Martin Everson, Laurence Lovat et al.

Purpose. Early squamous cell neoplasia (ESCN) in the oesophagus is a highly treatable condition. Lesions confined to the mucosal layer can be curatively treated endoscopically. We build a computer-assisted detection (CADe) system that can classify still images or video frames as normal or abnormal with high diagnostic accuracy. Methods. We present a new benchmark dataset containing 68K binary labeled frames extracted from 114 patient videos whose imaged areas have been resected and correlated to histopathology. Our novel convolutional network (CNN) architecture solves the binary classification task and explains what features of the input domain drive the decision-making process of the network. Results. The proposed method achieved an average accuracy of 91.7 % compared to the 94.7 % achieved by a group of 12 senior clinicians. Our novel network architecture produces deeply supervised activation heatmaps that suggest the network is looking at intrapapillary capillary loop (IPCL) patterns when predicting abnormality. Conclusion. We believe that this dataset and baseline method may serve as a reference for future benchmarks on both video frame classification and explainability in the context of ESCN detection. A future work path of high clinical relevance is the extension of the classification to ESCN types.

Luis C. Garcia-Peraza-Herrera, Lucas Fidon, Claudia D'Ettorre et al.

Producing manual, pixel-accurate, image segmentation labels is tedious and time-consuming. This is often a rate-limiting factor when large amounts of labeled images are required, such as for training deep convolutional networks for instrument-background segmentation in surgical scenes. No large datasets comparable to industry standards in the computer vision community are available for this task. To circumvent this problem, we propose to automate the creation of a realistic training dataset by exploiting techniques stemming from special effects and harnessing them to target training performance rather than visual appeal. Foreground data is captured by placing sample surgical instruments over a chroma key (a.k.a. green screen) in a controlled environment, thereby making extraction of the relevant image segment straightforward. Multiple lighting conditions and viewpoints can be captured and introduced in the simulation by moving the instruments and camera and modulating the light source. Background data is captured by collecting videos that do not contain instruments. In the absence of pre-existing instrument-free background videos, minimal labeling effort is required, just to select frames that do not contain surgical instruments from videos of surgical interventions freely available online. We compare different methods to blend instruments over tissue and propose a novel data augmentation approach that takes advantage of the plurality of options. We show that by training a vanilla U-Net on semi-synthetic data only and applying a simple post-processing, we are able to match the results of the same network trained on a publicly available manually labeled real dataset.

Luis C. Garcia-Peraza-Herrera, Martin Everson, Wenqi Li et al.

In this work, we have concentrated our efforts on the interpretability of classification results coming from a fully convolutional neural network. Motivated by the classification of oesophageal tissue for real-time detection of early squamous neoplasia, the most frequent kind of oesophageal cancer in Asia, we present a new dataset and a novel deep learning method that by means of deep supervision and a newly introduced concept, the embedded Class Activation Map (eCAM), focuses on the interpretability of results as a design constraint of a convolutional network. We present a new approach to visualise attention that aims to give some insights on those areas of the oesophageal tissue that lead a network to conclude that the images belong to a particular class and compare them with those visual features employed by clinicians to produce a clinical diagnosis. In comparison to a baseline method which does not feature deep supervision but provides attention by grafting Class Activation Maps, we improve the F1-score from 87.3% to 92.7% and provide more detailed attention maps.

Lucas Fidon, Wenqi Li, Luis C. Garcia-Peraza-Herrera et al.

The Dice score is widely used for binary segmentation due to its robustness to class imbalance. Soft generalisations of the Dice score allow it to be used as a loss function for training convolutional neural networks (CNN). Although CNNs trained using mean-class Dice score achieve state-of-the-art results on multi-class segmentation, this loss function does neither take advantage of inter-class relationships nor multi-scale information. We argue that an improved loss function should balance misclassifications to favour predictions that are semantically meaningful. This paper investigates these issues in the context of multi-class brain tumour segmentation. Our contribution is threefold. 1) We propose a semantically-informed generalisation of the Dice score for multi-class segmentation based on the Wasserstein distance on the probabilistic label space. 2) We propose a holistic CNN that embeds spatial information at multiple scales with deep supervision. 3) We show that the joint use of holistic CNNs and generalised Wasserstein Dice scores achieves segmentations that are more semantically meaningful for brain tumour segmentation.

Luis C. Garcia-Peraza-Herrera, Wenqi Li, Lucas Fidon et al.

Real-time tool segmentation from endoscopic videos is an essential part of many computer-assisted robotic surgical systems and of critical importance in robotic surgical data science. We propose two novel deep learning architectures for automatic segmentation of non-rigid surgical instruments. Both methods take advantage of automated deep-learning-based multi-scale feature extraction while trying to maintain an accurate segmentation quality at all resolutions. The two proposed methods encode the multi-scale constraint inside the network architecture. The first proposed architecture enforces it by cascaded aggregation of predictions and the second proposed network does it by means of a holistically-nested architecture where the loss at each scale is taken into account for the optimization process. As the proposed methods are for real-time semantic labeling, both present a reduced number of parameters. We propose the use of parametric rectified linear units for semantic labeling in these small architectures to increase the regularization ability of the design and maintain the segmentation accuracy without overfitting the training sets. We compare the proposed architectures against state-of-the-art fully convolutional networks. We validate our methods using existing benchmark datasets, including ex vivo cases with phantom tissue and different robotic surgical instruments present in the scene. Our results show a statistically significant improved Dice Similarity Coefficient over previous instrument segmentation methods. We analyze our design choices and discuss the key drivers for improving accuracy.

Lucas Fidon, Wenqi Li, Luis C. Garcia-Peraza-Herrera et al.

Brain tumour segmentation plays a key role in computer-assisted surgery. Deep neural networks have increased the accuracy of automatic segmentation significantly, however these models tend to generalise poorly to different imaging modalities than those for which they have been designed, thereby limiting their applications. For example, a network architecture initially designed for brain parcellation of monomodal T1 MRI can not be easily translated into an efficient tumour segmentation network that jointly utilises T1, T1c, Flair and T2 MRI. To tackle this, we propose a novel scalable multimodal deep learning architecture using new nested structures that explicitly leverage deep features within or across modalities. This aims at making the early layers of the architecture structured and sparse so that the final architecture becomes scalable to the number of modalities. We evaluate the scalable architecture for brain tumour segmentation and give evidence of its regularisation effect compared to the conventional concatenation approach.