International Journal of Computer Assisted Radiology and Surgery最新文献

{"title":"Analyzing pediatric forearm X-rays for fracture analysis using machine learning.","authors":"Van Lam, Abhijeet Parida, Sarah Dance, Sean Tabaie, Kevin Cleary, Syed Muhammad Anwar","doi":"10.1007/s11548-025-03485-z","DOIUrl":"https://doi.org/10.1007/s11548-025-03485-z","url":null,"abstract":"<p><strong>Purpose: </strong>Forearm fractures constitute a significant proportion of emergency department presentations in pediatric population. The treatment goal is to restore length and alignment between the distal and proximal bone fragments. While immobilization through splinting or casting is enough for non-displaced and minimally displaced fractures. However, moderately or severely displaced fractures often require reduction for realignment. However, appropriate treatment in current practices has challenges due to the lack of resources required for specialized pediatric care leading to delayed and unnecessary transfers between medical centers, which potentially create treatment complications and burdens. The purpose of this study is to build a machine learning model for analyzing forearm fractures to assist clinical centers that lack surgical expertise in pediatric orthopedics.</p><p><strong>Methods: </strong>X-ray scans from 1250 children were curated, preprocessed, and manually annotated at our clinical center. Several machine learning models were fine-tuned using a pretraining strategy leveraging self-supervised learning model with vision transformer backbone. We further employed strategies to identify the most important region related to fractures within the forearm X-ray. The model performance was evaluated with and without region of interest (ROI) detection to find an optimal model for forearm fracture analyses.</p><p><strong>Results: </strong>Our proposed strategy leverages self-supervised pretraining (without labels) followed by supervised fine-tuning (with labels). The fine-tuned model using regions cropped with ROI identification resulted in the highest classification performance with a true-positive rate (TPR) of 0.79, true-negative rate (TNR) of 0.74, AUROC of 0.81, and AUPR of 0.86 when evaluated on the testing data.</p><p><strong>Conclusion: </strong>The results showed the feasibility of using machine learning models in predicting the appropriate treatment for forearm fractures in pediatric cases. With further improvement, the algorithm could potentially be used as a tool to assist non-specialized orthopedic providers in diagnosing and providing treatment.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-24","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144700247","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"FLex: joint pose and dynamic radiance fields optimization for stereo endoscopic videos.","authors":"Florian Stilz, Mert Karaoglu, Felix Tristram, Nassir Navab, Benjamin Busam, Alexander Ladikos","doi":"10.1007/s11548-025-03446-6","DOIUrl":"https://doi.org/10.1007/s11548-025-03446-6","url":null,"abstract":"<p><strong>Purpose: </strong>Reconstruction of endoscopic scenes is crucial for various medical applications, from post-surgery analysis to educational training. However, existing methods are limited by static endoscopes, restricted deformation, or dependence on external tracking devices for camera pose information.</p><p><strong>Methods: </strong>We present flow-optimized local hexplanes (FLex), a novel approach addressing the challenges of a moving stereo endoscope in a highly dynamic environment. FLex implicitly separates the scene into multiple overlapping 4D neural radiance fields (NeRFs) and employs a progressive optimization scheme for joint reconstruction and camera pose estimation from scratch.</p><p><strong>Results: </strong>Tested on sequences of length up to 5000 frames, which is five times the length handled in the experiments of previous methods, this technique enhances usability substantially. It scales highly detailed reconstruction capabilities to significantly longer surgical videos, all without requiring external tracking information.</p><p><strong>Conclusion: </strong>Our proposed approach overcomes key limitations of existing methods by enabling accurate reconstruction and camera pose estimation for moving stereo endoscopes in challenging surgical scenes. FLex's advancements enhance the applicability of neural rendering techniques for medical applications, paving the way for improved surgical scene understanding. Code and data will be released on the project page: https://flexendo.github.io/.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-21","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144676460","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Influence of high-performance image-to-image translation networks on clinical visual assessment and outcome prediction: utilizing ultrasound to MRI translation in prostate cancer.","authors":"Mohammad R Salmanpour, Amin Mousavi, Yixi Xu, William B Weeks, Ilker Hacihaliloglu","doi":"10.1007/s11548-025-03481-3","DOIUrl":"https://doi.org/10.1007/s11548-025-03481-3","url":null,"abstract":"<p><strong>Purpose: </strong>Image-to-image (I2I) translation networks have emerged as promising tools for generating synthetic medical images; however, their clinical reliability and ability to preserve diagnostically relevant features remain underexplored. This study evaluates the performance of state-of-the-art 2D/3D I2I networks for converting ultrasound (US) images to synthetic MRI in prostate cancer (PCa) imaging. The novelty lies in combining radiomics, expert clinical evaluation, and classification performance to comprehensively benchmark these models for potential integration into real-world diagnostic workflows.</p><p><strong>Methods: </strong>A dataset of 794 PCa patients was analyzed using ten leading I2I networks to synthesize MRI from US input. Radiomics feature (RF) analysis was performed using Spearman correlation to assess whether high-performing networks (SSIM > 0.85) preserved quantitative imaging biomarkers. A qualitative evaluation by seven experienced physicians assessed the anatomical realism, presence of artifacts, and diagnostic interpretability of synthetic images. Additionally, classification tasks using synthetic images were conducted using two machine learning and one deep learning model to assess the practical diagnostic benefit.</p><p><strong>Results: </strong>Among all networks, 2D-Pix2Pix achieved the highest SSIM (0.855 ± 0.032). RF analysis showed that 76 out of 186 features were preserved post-translation, while the remainder were degraded or lost. Qualitative feedback revealed consistent issues with low-level feature preservation and artifact generation, particularly in lesion-rich regions. These evaluations were conducted to assess whether synthetic MRI retained clinically relevant patterns, supported expert interpretation, and improved diagnostic accuracy. Importantly, classification performance using synthetic MRI significantly exceeded that of US-based input, achieving average accuracy and AUC of ~ 0.93 ± 0.05.</p><p><strong>Conclusion: </strong>Although 2D-Pix2Pix showed the best overall performance in similarity and partial RF preservation, improvements are still required in lesion-level fidelity and artifact suppression. The combination of radiomics, qualitative, and classification analyses offered a holistic view of the current strengths and limitations of I2I models, supporting their potential in clinical applications pending further refinement and validation.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-19","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144668937","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Landmark-free automatic digital twin registration in robot-assisted partial nephrectomy using a generic end-to-end model.","authors":"Kilian Chandelon, Alice Pitout, Mathieu Souchaud, Julie Desternes, Gaëlle Margue, Julien Peyras, Nicolas Bourdel, Jean-Christophe Bernhard, Adrien Bartoli","doi":"10.1007/s11548-025-03473-3","DOIUrl":"https://doi.org/10.1007/s11548-025-03473-3","url":null,"abstract":"<p><strong>Purpose: </strong>Augmented Reality in Minimally Invasive Surgery has made tremendous progress in organs including the liver and the uterus. The core problem of Augmented Reality is registration, where a preoperative patient's geometric digital twin must be aligned with the image of the surgical camera. The case of the kidney is yet unresolved, owing to the absence of anatomical landmarks visible in both the patient's digital twin and the surgical images.</p><p><strong>Methods: </strong>We propose a landmark-free approach to registration, which is particularly well-adapted to the kidney. The approach involves a generic kidney model and an end-to-end neural network, which we train with a proposed dataset to regress the registration directly from a surgical RGB image.</p><p><strong>Results: </strong>Experimental evaluation across four clinical cases demonstrates strong concordance with expert-labelled registration, despite anatomical and motion variability. The proposed method achieved an average tumour contour alignment error of <math><mrow><mn>7.3</mn> <mo>±</mo> <mn>4.1</mn></mrow> </math> mm in <math><mrow><mn>9.4</mn> <mo>±</mo> <mn>0.2</mn></mrow> </math> ms.</p><p><strong>Conclusion: </strong>This landmark-free registration approach meets the accuracy, speed and resource constraints required in clinical practice, making it a promising tool for Augmented Reality-Assisted Partial Nephrectomy.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-17","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144661024","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Evaluating large language models on hospital health data for automated emergency triage.","authors":"Carlos Lafuente, Mehdi Rahim","doi":"10.1007/s11548-025-03475-1","DOIUrl":"https://doi.org/10.1007/s11548-025-03475-1","url":null,"abstract":"<p><strong>Purpose: </strong>Large language models (LLMs) have a significant potential in healthcare due to their ability to process unstructured text from electronic health records (EHRs) and to generate knowledge with few or no training. In this study, we investigate the effectiveness of LLMs for clinical decision support, specifically in the context of emergency department triage, where the volume of textual data is minimal compared to other scenarios such as making a clinical diagnosis.</p><p><strong>Methods: </strong>We benchmark LLMs with traditional machine learning (ML) approaches using the Emergency Severity Index (ESI) as the gold standard criteria of triage. The benchmark includes general purpose, specialised, and fine-tuned LLMs. All models are prompted to predict ESI score from a EHRs. We use a balanced subset (n = 1000) from MIMIC-IV-ED, a large database containing records of admissions to the emergency department of Beth Israel Deaconess Medical Center.</p><p><strong>Results: </strong>Our findings show that the best-performing models have an average F1-score below 0.60. Also, while zero-shot and fine-tuned LLMs can outperform standard ML models, their performance is surpassed by ML models augmented with features derived from LLMs or knowledge graphs.</p><p><strong>Conclusion: </strong>LLMs show value for clinical decision support in scenarios with limited textual data, such as emergency department triage. The study advocates for integrating LLM knowledge representation to improve existing ML models rather than using LLMs in isolation, suggesting this as a more promising approach to enhance the accuracy of automated triage systems.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-16","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144644112","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Efficient needle guidance: multi-camera augmented reality navigation without patient-specific calibration.","authors":"Yizhi Wei, Bingyu Huang, Bolin Zhao, Zhengyu Lin, Steven Zhiying Zhou","doi":"10.1007/s11548-025-03477-z","DOIUrl":"https://doi.org/10.1007/s11548-025-03477-z","url":null,"abstract":"<p><strong>Purpose: </strong>Augmented reality (AR) technology holds significant promise for enhancing surgical navigation in needle-based procedures such as biopsies and ablations. However, most existing AR systems rely on patient-specific markers, which disrupt clinical workflows and require time-consuming preoperative calibrations, thereby hindering operational efficiency and precision.</p><p><strong>Methods: </strong>We developed a novel multi-camera AR navigation system that eliminates the need for patient-specific markers by utilizing ceiling-mounted markers mapped to fixed medical imaging devices. A hierarchical optimization framework integrates both marker mapping and multi-camera calibration. Deep learning techniques are employed to enhance marker detection and registration accuracy. Additionally, a vision-based pose compensation method is implemented to mitigate errors caused by patient movement, improving overall positional accuracy.</p><p><strong>Results: </strong>Validation through phantom experiments and simulated clinical scenarios demonstrated an average puncture accuracy of 3.72 ± 1.21 mm. The system reduced needle placement time by 20 s compared to traditional marker-based methods. It also effectively corrected errors induced by patient movement, with a mean positional error of 0.38 pixels and an angular deviation of 0.51 <math><mmultiscripts><mrow></mrow> <mrow></mrow> <mo>∘</mo></mmultiscripts> </math> . These results highlight the system's precision, adaptability, and reliability in realistic surgical conditions.</p><p><strong>Conclusion: </strong>This marker-free AR guidance system significantly streamlines surgical workflows while enhancing needle navigation accuracy. Its simplicity, cost-effectiveness, and adaptability make it an ideal solution for both high- and low-resource clinical environments, offering the potential for improved precision, reduced procedural time, and better patient outcomes.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144621053","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Assessment of cognitive load in the context of neurosurgery.","authors":"Daniel A Di Giovanni, M Kersten-Oertel, S Drouin, D L Collins","doi":"10.1007/s11548-025-03478-y","DOIUrl":"https://doi.org/10.1007/s11548-025-03478-y","url":null,"abstract":"<p><strong>Purpose: </strong>Image-guided neurosurgery demands precise depth perception to minimize cognitive burden during intricate navigational tasks. Existing evaluation methods rely heavily on subjective user feedback, which can be biased and inconsistent. This study uses a physiological measure via electroencephalography (EEG), to quantify cognitive load when using novel dynamic depth-cue visualizations. By comparing dynamic versus static rendering techniques, we aim to establish an objective framework for assessing and validating visualization strategies beyond traditional performance metrics.</p><p><strong>Methods: </strong>Twenty participants (experts in brain imaging) navigated to specified targets within a computed tomography angiography (CTA) volume using a tracked 3D pointer. We implemented three visualization methods (shading, ChromaDepth, aerial perspective) in both static and dynamic modes, randomized across 80 trials per subject. Continuous EEG was recorded via a Muse headband; raw signals were preprocessed and theta-band (4-7 Hz) power extracted for each trial. A two-way repeated measures ANOVA assessed the effects of visualization type and dynamic interaction on theta power.</p><p><strong>Results: </strong>Dynamic visualization conditions yielded lower mean theta-band power compared to static conditions (Δ = 0.057 V2/Hz; F (1,19) = 6.00, p = 0.024), indicating reduced neural markers of cognitive load. No significant main effect was observed across visualization methods, nor their interaction with dynamic mode. These findings suggest that real-time feedback from pointer-driven interactions may alleviate mental effort regardless of the specific depth cue employed.</p><p><strong>Conclusion: </strong>Our exploratory results demonstrate the feasibility of using consumer-grade EEG to provide objective insights into cognitive load for surgical visualization techniques. Although limited by non-surgeon participants, the observed theta-power reductions under dynamic conditions support further investigation. Future work should correlate EEG-derived load measures with performance outcomes, involve practising neurosurgeons, and leverage high-density EEG or AI-driven adaptive visualization to refine and validate these preliminary findings.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144621079","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"Comparison of the accuracy of different slot properties of 3D-printed cutting guides for raising free fibular flaps using saw or piezoelectric instruments: an in vitro study.","authors":"Britta Maria Lohn, Stefan Raith, Mark Ooms, Philipp Winnand, Frank Hölzle, Ali Modabber","doi":"10.1007/s11548-025-03474-2","DOIUrl":"https://doi.org/10.1007/s11548-025-03474-2","url":null,"abstract":"<p><strong>Purpose: </strong>The free fibular flap (FFF) is a standard procedure for the oral rehabilitation of segmental bone defects in the mandible caused by diseases such as malignant processes, osteonecrosis, or trauma. Digital guides and computer-assisted surgery (CAS) can improve precision and reduce the time and cost of surgery. This study evaluates how different designs of slot cutting guides, guiding heights, and cutting instruments affect surgical accuracy during mandibular reconstruction.</p><p><strong>Methods: </strong>Ninety model operations in a three-part fibular transplant for mandibular reconstruction were conducted according to digital planning with three guide designs (standard, flange, and anatomical slots), three guide heights (1 mm, 2 mm, 3 mm), and two osteotomy instruments (piezoelectric instrument and saw). The cut segments were digitized using computed tomography and digitally evaluated to assess surgical accuracy.</p><p><strong>Results: </strong>For vestibular and lingual segment length, the anatomical slot and the flange appear to be the most accurate, with the flange slightly under-contoured vestibularly and the standard slot over-contoured lingually and vestibularly (p < 0.001). There were only minor differences between the use of saw and piezoelectric instrument for lingual (p = 0.005) and vestibular (p < 0.001) length and proximal angle (p = 0.014). The U-distance after global reconstruction for flanges resulted in a median deviation of 0.0468 mm (IQR 8.15), but was not significant (p = 0.067).</p><p><strong>Conclusion: </strong>Anatomical slots and flanges are recommended for osteotomy, with guiding effects relying on both haptic and visual control. Unilateral guided flanges also work accurately at high guidance heights. The results of piezoelectric instrument (PI) and saw showed comparable results in the assessment of individual segments and U-reconstruction in this in vitro study without soft tissue, so that the final decision is left to the expertise of the surgeons.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-12","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144621052","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"GPU-accelerated deformation mapping in hybrid organ models for real-time simulation.","authors":"Rintaro Miyazaki, Yuichiro Hayashi, Masahiro Oda, Kensaku Mori","doi":"10.1007/s11548-025-03377-2","DOIUrl":"https://doi.org/10.1007/s11548-025-03377-2","url":null,"abstract":"<p><strong>Purpose: </strong>Surgical simulation is expected to be an effective way for physicians and medical students to learn surgical skills. To achieve real-time deformation of soft tissues with high visual quality, multiple resolution and adaptive mesh refinement models have been introduced. However, those models require additional processing time to map the deformation results of the deformed lattice to a polygon model. In this study, we propose a method to accelerate this process using vertex shaders on GPU and investigate its performance.</p><p><strong>Methods: </strong>A hierarchical octree cube structure is generated from a high-resolution organ polygon model. The entire organ model is divided into pieces according to the cube structure. In a simulation, vertex coordinates of the organ model pieces are obtained by trilinear interpolation of the cube's 8 vertex coordinates. This process is described in a shader program, and organ model vertices are processed in the rendering pipeline for acceleration.</p><p><strong>Results: </strong>For a constant number of processing cubes, the CPU-based processing time increased linearly with the total number of organ model vertices, and the GPU-based time was nearly constant. On the other hand, for a constant number of model vertices, the GPU-based time increased linearly with the number of surface cubes. These linearities determine a condition that the GPU-based implementation is faster in the same frame time.</p><p><strong>Conclusion: </strong>We implemented octree cube deformation mapping using vertex shaders and confirmed its performance. The experimental results showed that the GPU can accelerate the mapping process in high-resolution organ models with a large number of vertices.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144576905","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}

{"title":"eNCApsulate: neural cellular automata for precision diagnosis on capsule endoscopes.","authors":"Henry John Krumb, Anirban Mukhopadhyay","doi":"10.1007/s11548-025-03425-x","DOIUrl":"https://doi.org/10.1007/s11548-025-03425-x","url":null,"abstract":"<p><strong>Purpose: </strong>Wireless capsule endoscopy (WCE) is a noninvasive imaging method for the entire gastrointestinal tract and is a pain-free alternative to traditional endoscopy. It generates extensive video data that requires significant review time, and localizing the capsule after ingestion is a challenge. Techniques like bleeding detection and depth estimation can help with localization of pathologies, but deep learning models are typically too large to run directly on the capsule.</p><p><strong>Methods: </strong>Neural cellular automata (NCAs) for bleeding segmentation and depth estimation are trained on capsule endoscopic images. For monocular depth estimation, we distill a large foundation model into the lean NCA architecture, by treating the outputs of the foundation model as pseudo-ground truth. We then port the trained NCAs to the ESP32 microcontroller, enabling efficient image processing on hardware as small as a camera capsule.</p><p><strong>Results: </strong>NCAs are more accurate (Dice) than other portable segmentation models, while requiring more than 100x fewer parameters stored in memory than other small-scale models. The visual results of NCAs depth estimation look convincing and in some cases beat the realism and detail of the pseudo-ground truth. Runtime optimizations on the ESP32-S3 accelerate the average inference speed significantly, by more than factor 3.</p><p><strong>Conclusion: </strong>With several algorithmic adjustments and distillation, it is possible to eNCApsulate NCA models into microcontrollers that fit into wireless capsule endoscopes. This is the first work that enables reliable bleeding segmentation and depth estimation on a miniaturized device, paving the way for precise diagnosis combined with visual odometry as a means of precise localization of the capsule-on the capsule.</p>","PeriodicalId":51251,"journal":{"name":"International Journal of Computer Assisted Radiology and Surgery","volume":" ","pages":""},"PeriodicalIF":2.3,"publicationDate":"2025-07-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144565483","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":3,"RegionCategory":"医学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}