SurgX in Surgical AI: Explainability & 3D Reconstruction

Updated 6 July 2026

SurgX for phase recognition leverages neuron-concept associations with surgery-specific concept sets to provide explainable, temporally-aware interpretations of surgical workflows, achieving improved alignment and interpretability metrics on Cholec80.
SurgX as IXGS employs R²-Gaussian splatting and radiographic standardization to reconstruct anatomically meaningful 3D lumbar-spine volumes from sparse, arbitrarily posed X-rays, enhancing surgical navigation.
The term SurgX functions as an overloaded label for distinct AI approaches in surgery, highlighting its dual role in both explainability for surgical phase analysis and intraoperative 3D reconstruction.

Searching arXiv for the provided SurgX-related papers and nearby terminology to ground the article. SurgX is a label used in recent surgical AI literature for more than one technically distinct system. In one usage, it denotes a concept-based explanation framework for surgical phase recognition that associates neurons with clinically meaningful surgical concepts and explains predictions through the concepts of highly contributing neurons (Kim et al., 21 Jul 2025). In another, the term is used in the context of intraoperative 3D reconstruction, where it refers to an instance-based framework formally presented as IXGS for reconstructing anatomically meaningful lumbar-spine volumes from sparse, arbitrarily posed real X-rays (Jecklin et al., 20 Apr 2025). This suggests a shared emphasis on clinically grounded surgical AI, but the two systems address different tasks, modalities, and deployment scenarios.

1. Terminological scope and disambiguation

A common source of confusion is that SurgX is not a single universally standardized platform in the available literature. Rather, the name appears in at least two distinct senses, each tied to a different paper and problem formulation (Kim et al., 21 Jul 2025, Jecklin et al., 20 Apr 2025).

Usage of “SurgX”	Technical problem	Core mechanism
SurgX for surgical phase recognition	Explainability for video-based phase prediction	Neuron-concept association with SurgVLP embeddings
SurgX as IXGS	Intraoperative 3D reconstruction from sparse X-rays	$R^2$ -Gaussian splatting with arbitrary poses and radiographic standardization

The first usage is centered on laparoscopic workflow analysis. The second is centered on navigation-oriented spine imaging. The term therefore functions as an overloaded label rather than a single research line. Any precise discussion of SurgX requires identifying which of these two formulations is intended.

2. SurgX for explainable surgical phase recognition

In "SurgX: Neuron-Concept Association for Explainable Surgical Phase Recognition" (Kim et al., 21 Jul 2025), SurgX is a concept-based explainability framework for surgical phase recognition. Its stated motivation is that surgical phase recognition is crucial for surgical workflow analysis, surgical monitoring, skill assessment, and workflow optimization, yet modern deep learning models for this task, especially TCNs and Transformers, remain black boxes. The paper treats this opacity as problematic for trust, debugging, and regulatory transparency, particularly because surgical video models depend on temporal context rather than isolated frames.

The framework is evaluated on two target models: TeCNO, a TCN-based surgical phase recognition model, and Causal ASFormer, a Transformer-based temporal action segmentation model modified to be causal and to use LoViT features as spatial features. The dataset is Cholec80, comprising 80 laparoscopic cholecystectomy videos annotated with surgical phases, split into 40 training videos and 40 test videos. The training set is also used as the probing set for neuron representation selection (Kim et al., 21 Jul 2025).

A central design choice is the use of surgery-specific concept sets rather than generic visual concepts. The paper defines three concept sets. CholecT45-W is built from action triplet labels in CholecT45 and contains 30 words such as grasper, clipping, and gallbladder. CholecT45-S is also built from CholecT45 action triplets and contains 100 sentences generated by a prompting approach, including examples such as “I use a hook to dissect the cystic plate.” ChoLec-270 is the largest concept set, with 270 concepts collected from 11 cholecystectomy lecture videos from WebSurg and 4 cholecystectomy-related articles. The paper reports that this richer and more specialized concept inventory gives the best quantitative interpretability results (Kim et al., 21 Jul 2025).

3. Neuron-concept association methodology

SurgX for phase recognition consists of three stages: construction of a surgical concept set, selection of representative example sequences for each neuron, and annotation of neurons with concepts followed by prediction explanation (Kim et al., 21 Jul 2025). The representation of a neuron is explicitly temporal. Rather than selecting a single frame only, the method first identifies highly activated frames after ReLU and then constructs a sequence by including the selected frame plus $n$ previous frames sampled with a dilation rate $m$ . The paper argues that this is necessary because, in phase recognition, a neuron’s meaning may depend on surrounding frames rather than a single instant.

For neuron annotation, text features from the concept set and visual features from the representative sequences are extracted with SurgVLP, a surgical vision-LLM chosen to reduce embedding mismatch between surgical video and concept text. For neuron $n$ and concept $c$ , the concept score is defined as

$s_{n,c} = \frac{1}{E} \sum_{e=1}^{E} \cos\bigl(v_{n,e}, t_c\bigr),$

where $E$ is the number of frames in the representative example set, $v_{n,e}$ is the visual feature of the $e$ -th example frame, and $t_c$ is the text feature of concept $n$ 0. A concept is annotated to a neuron when

$n$ 1

Prediction explanation is based on highly contributing neurons in the penultimate layer. The paper defines neuron contribution counterfactually as

$n$ 2

Concepts associated with these highly contributing neurons are then presented as the explanation for the predicted phase. The resulting explanations are conceptual rather than pixel-level, with examples such as “insert a port,” “pushed into the port,” “hepatocystic triangle,” and “cystic artery is isolated between clips” (Kim et al., 21 Jul 2025).

4. Empirical behavior, ablations, and failure analysis

The paper evaluates interpretability with two metrics: Concept Alignment Score, computed on the final layer by comparing neuron concepts with the true phase expressed in word or sentence form, and Prediction Interpretability Score, computed on the penultimate layer by comparing the predicted phase with the concepts annotated to the most contributing neurons (Kim et al., 21 Jul 2025). Across the tested concept sets, ChoLec-270 performs best.

Concept set	Concept Alignment Avg	Prediction Interpretability Avg
CholecT45-W	0.3880	0.5539
CholecT45-S	0.3769	0.5122
ChoLec-270	0.4475	0.5992

The representative-frame and sequence-selection ablations further refine the design. Video-wise Threshold gives the best overall concept alignment among frame-selection strategies. Sequence-based neuron representation outperforms single-frame representation, and the best concept alignment is achieved by Dilated-Sequence (5), with Avg Concept Alignment of 0.4475. The best Avg Prediction Interpretability in one metric is achieved by Dilated-Sequence (10), with 0.6202. The paper nevertheless identifies the dilated sequence with 5-second interval as the best overall setting (Kim et al., 21 Jul 2025).

Qualitative case studies show that correct predictions can depend on meaningful temporal cues even when the current frame is visually ambiguous. For the Preparation phase, neurons associated with “Insert a port” and “Pushed into the port” contribute strongly, even if the current frame does not show a port. For Clipping and Cutting, neurons associated with “hepatocystic triangle” and “cystic artery is isolated between clips” help prediction even when the frame is dark. The framework also exposes systematic error modes. When the true phase is Gallbladder Dissection, mispredictions as Clipping and Cutting often involve neurons associated with “cystic artery is isolated between clips”: 88.22% of mispredicted cases involve these neurons, whereas in correctly predicted cases 92.88% do not involve them (Kim et al., 21 Jul 2025). This makes the framework useful not only for post hoc explanation but also for diagnosing recurrent misclassification patterns.

The paper’s limitations are mostly implicit. It is evaluated only on Cholec80 and cholecystectomy-related concepts; the concept sets are curated; explanations depend on SurgVLP; and validation is limited to TeCNO and Causal ASFormer. A plausible implication is that extension to other procedures would require new concept inventories and renewed analysis of concept-neuron alignment.

5. SurgX as IXGS: intraoperative 3D reconstruction from sparse X-rays

In "IXGS-Intraoperative 3D Reconstruction from Sparse, Arbitrarily Posed Real X-rays" (Jecklin et al., 20 Apr 2025), SurgX refers to the authors’ instance-based intraoperative 3D reconstruction framework, formally presented as IXGS. Here the problem is not explainability but reconstruction of anatomically meaningful 3D lumbar-spine volumes from sparse, arbitrarily posed real fluoroscopic or X-ray images for surgical navigation, especially pedicle screw placement. The clinical motivation is that CBCT provides excellent guidance but increases radiation dose, acquisition complexity, and workflow burden, whereas ordinary 2D fluoroscopy is low-dose and widely available but provides limited perspective on 3D anatomy.

IXGS adapts $n$ 3-Gaussian splatting to a setting in which views are sparse, arbitrarily posed, and inconsistent in appearance. The 3D anatomy is represented as a set of Gaussian kernels

$n$ 4

with individual kernels

$n$ 5

where $n$ 6 is the kernel center, $n$ 7 its covariance, and $n$ 8 its central density. Because X-ray imaging is density-based, the method does not model color. The framework generalizes the rectified projection mechanism of $n$ 9-Gaussian to arbitrary C-arm poses and adds an anatomy-guided radiographic standardization stage using a Pix2Pix style-transfer network trained on paired real X-rays and synthetic DRRs. The purpose of this standardization is to reduce inter-view appearance variation, improve consistency of bone contrast, and stabilize Gaussian-splatting optimization (Jecklin et al., 20 Apr 2025).

The preprocessing and optimization pipeline is tailored to real intraoperative geometry. Images are cropped around the calibrated principal point because the C-arm principal point is often off-center, and the target reconstruction volume is normalized to $m$ 0. Since the data are arbitrary-view real images, the method cannot rely on FDK initialization, so the 3D Gaussian kernels are randomly initialized inside the reconstruction volume. Optimization compares rendered projections with input views using an objective that combines L1 loss, SSIM, and total variation, with adaptive density control inherited from Gaussian splatting. After optimization, the learned Gaussian set is converted to a voxel volume with a differentiable voxelizer, thresholded, and cropped to focus on the spine; the threshold is empirically set around the 80th percentile (Jecklin et al., 20 Apr 2025).

A notable property of IXGS is that no anatomy-specific pretraining is required for the Gaussian reconstruction model. The trade-off is computational cost and the need for more views. The reconstruction model is trained for 30k iterations on an NVIDIA A100 40 GB GPU, while the Pix2Pix standardization network is trained for 200 epochs with batch size 1, learning rate $m$ 1, and Adam with $m$ 2 (Jecklin et al., 20 Apr 2025).

The study uses an ex-vivo paired real/synthetic lumbar-spine dataset comprising six ex-vivo human specimens with calibrated real X-rays and corresponding DRRs from CT. For each specimen, 50 X-rays are randomly sampled for training, yielding 300 training images total, and the remaining views form a 233-image test set. Views include AP, lateral, oblique, and other off-axis angles acquired with a clinical mobile C-arm over wide angular ranges. Experiments compare synthetic circular views, synthetic arbitrary DRRs, raw real X-rays, and style-transferred real X-rays, while varying the number of input views from 5 to 50 in steps of 5 (Jecklin et al., 20 Apr 2025).

The quantitative results show a consistent ordering. For 50 views, the circular synthetic baseline reaches 39.19 dB PSNR and 0.970 SSIM; synthetic arbitrary DRRs reach 29.45 and 0.890; raw real X-rays reach 23.22 and 0.760; and style-transferred real X-rays improve this to 25.73 and 0.790. For 25 views, the corresponding numbers are 33.51 and 0.940 for circular synthetic, 26.62 and 0.830 for synthetic arbitrary, 20.52 and 0.720 for raw real, and 24.17 and 0.750 for style-transferred real X-rays (Jecklin et al., 20 Apr 2025). The paper emphasizes that standardization improves over raw real inputs without closing the full gap to idealized synthetic data.

Clinical utility is assessed directly by an experienced orthopedic surgeon using a 4-point Likert scale. Direct 3D volume renderings are generally “Poor” or “Unusable” below about 15 views, become sometimes “Acceptable” around 20 views, and reach “Very Good” for some cases at about 30 views. Slice views show “Acceptable” beginning around 15 views and generally improve with more views; at 50 views, most slice-based assessments are “Very Good.” The surgeon preferred slice browsing over direct volume rendering because scrolling through slices gave better spatial perception for navigation planning. The paper reports reconstruction time for 50 style-transferred images of about 13 min 33 s, and it notes residual “cloudy” artifacts in sparsely constrained regions caused by random initialization and incomplete coverage (Jecklin et al., 20 Apr 2025).

6. Relation to adjacent surgical AI systems

The two principal meanings of SurgX sit within a wider surgical AI landscape that includes dynamic scene reconstruction, perioperative multi-agent reasoning, voice-directed operating-room interaction, and AI-XR operating environments. In endoscopic 3D reconstruction, "SurgicalGS: Dynamic 3D Gaussian Splatting for Accurate Robotic-Assisted Surgical Scene Reconstruction" addresses fast, dynamic, and geometrically accurate 3D reconstruction from endoscopic video, arguing that inverse-depth supervision compresses depth variations and proposing dense initialization from depth priors, dynamic deformation modeling, and normalized edge-aware depth supervision (Chen et al., 2024). This is adjacent to IXGS in its use of Gaussian primitives, but it targets dynamic endoscopic tissue scenes rather than sparse X-ray reconstruction.

In perioperative decision support, "SURGENT: A Surgical Multi-Agent Assistance System Across the Perioperative Workflow" defines a planner-guided, memory-augmented multi-agent system spanning case analysis, surgical plan simulation, safety monitoring, complication risk assessment, and rehabilitation guidance (Shi et al., 28 May 2026). Its architecture combines a Tree-of-Thought planner, dual-memory design, departmental collaboration agents, aggregation, reflection, and human review. This line of work is orthogonal to both meanings of SurgX, but it reflects the same movement toward clinically grounded, auditable surgical AI.

In robotic operating-room interaction, the "Surgical Agent Orchestration Platform for Voice-directed Patient Data Interaction" presents SAOP as a hierarchical multi-agent system for da Vinci surgery, enabling voice-driven retrieval of patient information, manipulation of CT views, and navigation of 3D anatomical overlays on surgical video (Park et al., 10 Nov 2025). Its emphasis on workflow continuity, multimodal interaction, and robustness to speech-recognition errors provides another nearby point in the design space of surgical AI systems intended for active procedural use.

A broader conceptual context is provided by "Can We Revitalize Interventional Healthcare with AI-XR Surgical Metaverses?" which frames AI-XR surgery as a workflow spanning preoperative planning, consultation, intraoperative guidance, training, telementoring, and telesurgery, while insisting that such systems must be secure, robust, trustworthy, and clinically dependable (Qayyum et al., 2023). Its proof-of-concept “immersive surgical attack” on incision point localization shows that even minute manipulations of a digital twin can shift planned incision locations in clinically meaningful ways. This security perspective is especially relevant to any future extension of SurgX-style systems into immersive or operative environments.

Taken together, these neighboring systems clarify that SurgX is best understood not as a single canonical product but as a name attached to two distinct surgical AI formulations: one oriented toward explanation of temporal video recognition, the other toward patient-specific reconstruction from sparse intraoperative imaging. The former treats interpretability as the central problem; the latter treats anatomically useful reconstruction without pretraining as the central problem. Their coexistence under the same label reflects the breadth of current surgical AI research rather than a unified technical lineage.