Spatiotemporal Recognition of Abdominal Wall Layers During Optical Trocar Entry

• Trocar insertion can cause serious vascular or visceral injury, while comparisons of entry techniques have not established a universally superior approach for preventing major injuries [1–5].The rarity of these events and underpowered comparative studies contribute to the uncertainty.Source: §1 · p. 1
• Multiple tissues may appear in one view, so recognition must capture their spatial distribution and use neighboring-region and sequence information beyond local appearance.Glare, similar coloration across fat compartments, and thin layers further complicate recognition.Source: §1 · p. 1
• The introduction motivates assessing class-specific performance, spatial consistency, onset timing, and probability quality because accuracy alone cannot distinguish noise reduction from excessive smoothing [11,12].Source: §1 · p. 1
• The study evaluates nine-class recognition across 28 sectors using recording-grouped cross-validation and clinician-derived primary targets, while noting that its temporal pipeline uses future observations.Source: §1 · p. 2
• Macro-F1 improved from 0.440 to 0.507 with the integrated approach, with gains across the five folds.The introduction presents this as part of a broader evaluation of classification, probability, spatial, and temporal outcomes.Source: §1 · p. 2

2.1 Entry Safety and Training

• Entry-safety studies span adverse-event reports and strategy reviews, with syntheses finding insufficient evidence of superiority on major injuries [4,5]; training research addresses technical performance and skill transfer [32–39].Video-based assessment does not by itself establish educational effectiveness or improved surgical safety [42–47].Source: §2.1 · p. 2

2.2 Surgical Video Recognition

• Surgical video research covers workflow, instruments, and anatomical tasks [6–8,53–68], whereas trocar-entry recognition must distinguish multiple simultaneously visible tissues [56,57].Existing datasets and acquisition settings differ from the abdominal-wall sectors examined here.Source: §2.2 · p. 2

2.3 Temporal Models and Efficient Features

• HMMs, semi-Markov models, and CRFs provide structured sequence constraints [69–71]; this framework combines learned posterior smoothing with spatial and anatomical constraints, without evaluating duration models or neural sequence architectures.Efficient neural representations motivate future comparisons but were not evaluated in this study.Source: §2.3 · p. 2

2.4 Annotation and Predictive Reliability

• Evaluation guidance motivates preserving annotation provenance and reporting discrimination, probability quality, and temporal errors separately [78–86]; this study integrates established modeling ideas for tissue recognition, not injury prevention.Figure 1 depicts temporal smoothing, within-frame MRF regularization, adjacent-state HMM transitions, and blending with the original emission probabilities.Source: §2.4 · pp. 2–3, Figure 1 · p. 3

3.1 Study Design and Video Dataset

• The retrospective framework classified 28 sectors per circular view into tissue classes and compared methods using five-fold cross-validation grouped by recording.It processed 71 recordings, aligned annotations to video frames, and required at least two annotated keyframes per recording.Source: §3.1 · p. 3

3.2 Tissue Vocabulary and Annotation Paradigms

• Nine tissue classes were ordered anatomically, and dense and original annotation paradigms were combined with transferred labels for training while primary evaluation used clinician-drawn pixels.Primary classification used clinician-only targets; secondary classification used all targets.Source: §3.2 · p. 3, §3.1 · p. 3

3.3 Deterministic Annotation Transfer

• Transfer left keyframes unchanged when clinician masks covered at least 65% of the field of view and otherwise permitted restricted peripheral filling when specified outer-wall structures were present.Only subcutaneous fat was newly introduced; validation compared transferred labels with dense clinician annotations using per-class pixel IoU.Source: §3.3 · pp. 3–4

3.4 Video Preparation and Source-Frame Alignment

• Videos were restricted to annotated spans, center-cropped to 1080 × 1080 pixels, resized to 512 × 512, and temporally aligned before feature extraction.Source: §3.4 · p. 4

3.5 Field-of-View Detection and Polar Geometry

• The circular field of view was estimated from the first source frame, then divided into four equal-width rings and 28 sectors, with four inner and eight outer angular subdivisions per ring.Neighbor context used inward, outward, and angular sectors; where radial neighbors did not exist, the sector itself supplied context features.Source: §3.5 · p. 4

3.6 Sector Targets and Primary Evaluation Labels

• Targets required labeled pixels to cover at least one-quarter of a sector; unlabeled sectors were excluded from scoring but could contribute context features, and primary evaluation used clinician-drawn pixels.For transferred recordings, added subcutaneous fat was excluded from primary targets, although eligible training targets included transferred labels.Source: §3.6 · p. 4

3.7 Color Features and Specular-Glare Handling

• Each sector used a 173-dimensional descriptor combining Lab and HSV color features; glare was excluded from histograms and moments but retained as a fraction, and hue was rotated to keep red contiguous.Framewise white balancing over nonglare field-of-view pixels was intended to reduce illumination and color-balance variation.Source: §3.7 · pp. 4–5

3.8 Neighbor Context and Positional Encoding

• The 245-dimensional representation combines the sector descriptor with five 13-dimensional context blocks and a seven-dimensional vector encoding ring, radial, and angular position.Context blocks represent inner and outer neighbors, angular-neighbor average, frame mean, and the sector–frame-mean difference.Source: §3.8 · p. 5

3.9 Emission Classifier and Video-Grouped Validation

• A class-balanced gradient-boosting emission model uses fixed settings and no preceding or subsequent frames, with five-fold validation grouped by recording.The resulting class-balanced posterior is treated as an approximation to a scaled class-conditional likelihood for later inference.Source: §3.9 · pp. 5–6

3.10 Learned Temporal Smoothing and Cross-Fitting

• The temporal smoother combines emission probabilities over short and longer past-and-future windows, while cross-fitting keeps each training recording out of its first-stage predictions.On the 60-fps grid, windows span approximately the preceding and following one second and one-to-five seconds; the reported full pipeline is bidirectional.Source: §3.10 · p. 6

3.11 Within-Frame Spatial Regularization

• The MRF couples adjacent-sector classes using pointwise mutual information estimated from training-keyframe co-label counts with add-one smoothing.Mean-field inference starts from the incoming posterior, which is the smoother output in the complete pipeline; pairwise counts are estimated from training folds only.Source: §3.11 · p. 6
• Alternative spatial comparisons included a depth-and-ring hierarchical model, learned spatial stacking, and combinations with a banded HMM or cross-fitted decision biases.The spatial-stacking model used 61 features derived from sector, neighboring-sector, and frame-mean posteriors.Source: §3.11 · p. 6

3.12 Anatomically Ordered Hidden Markov Models

• The independent nine-state HMM orders states by tissue code and permits persistence, one-state advancement, or one-state retreat, but no state can be skipped in one grid step.Transition rates are estimated separately for each training fold; log-space forward–backward inference maintains numerical stability.Source: §3.12 · pp. 6–7

3.13 Posterior Blending and Calibration Diagnostics

• The final sequence applies emission classification, learned temporal smoothing, MRF regularization, HMM smoothing, then geometric blending with the original emission posterior.The blend weight was 0.75; the raw classifier's class ranking is retained alongside the regularized probabilities.Source: §3.13 · p. 7
• Temperature scaling is evaluated separately as a log-loss diagnostic and preserves the maximum-probability class, leaving accuracy and F1 unchanged.Fold-specific temperatures are fitted using other folds' out-of-fold outputs; these diagnostic results are separate from the main unadjusted probability metrics.Source: §3.13 · p. 7

3.14 Classification and Spatial-Consistency Endpoints

• Primary classification evaluation scores clinician-only labeled sectors on annotated keyframes, with accuracy, macro-F1, and macro AUROC as headline endpoints.Spatial consistency is the proportion of eligible frames whose annotated-sector tissue-code depth span is at least three.Source: §3.14 · p. 7

3.15 Temporal Coherence and Peritoneal-Onset Endpoints

• Temporal evaluation measures label changes per sector-minute and onset error relative to the interval between the last shallower and first peritoneum-or-deeper clinician keyframes.Reported onset measures include the proportion with error at most two seconds, premature and missed onsets, and median interval error.Source: §3.15 · p. 7

4.1 Dataset Assembly and Preprocessing Verification

• The final dataset comprised 71 videos, 1,516 keyframes, and 29,151 clinician-only sectors; all nine tissue classes were retained despite sparse bowel and preperitoneal-fat sectors.Those two classes accounted for 503 and 537 clinician-only sectors, respectively.Source: §4.1 · pp. 7–8

4.2 Annotation-Transfer Validation

• The adopted transfer procedure achieved mean per-class pixel IoU of 0.873; broader filling raised isolated-label correctness to 89% but reduced overall IoU to 0.792.About 78% of subcutaneous-fat sectors originated from filling, and primary subcutaneous-fat evaluation used densely annotated recordings only.Source: §4.2 · p. 8

4.3 Primary Classification Results

• Adding learned smoothing to the MRF–HMM–blend pipeline raised macro-F1 from 0.474 ± 0.051 to 0.507 ± 0.049 on clinician-only targets.The reported means also increased by 0.018 for accuracy and 0.014 for AUROC.Source: §4.3 · p. 8
• Spatial regularization alone produced macro-F1 of 0.462 ± 0.044, while the strict HMM produced macro-F1 of 0.429 ± 0.041.The strict HMM’s paired F1 change was −0.011 ± 0.023; the MRF–HMM combination had a positive paired change in every fold.Source: §4.3 · p. 8

4.4 Secondary Evaluation Using All Labels

• The complete pipeline achieved 0.682 accuracy, 0.495 macro-F1, and 0.886 AUROC on all labeled sectors, with paired macro-F1 gains in every fold.The full pipeline had the highest macro-F1, macro AUROC, and balanced accuracy in this evaluation.Source: §4.4 · p. 8, Table 4 · p. 9

4.5 Within-Frame Tissue-Depth Spread

• The full pipeline had depth spans of at least three tissue codes in 26% of eligible keyframes, compared with 22% for clinician annotations and 53% for emissions.MRF regularization alone reduced the proportion to 21%.Source: §4.5 · p. 8

4.6 Temporal Coherence and Onset Recognition

• Emissions changed labels 66.2 ± 3.7 times per sector-minute, compared with 43.0 ± 4.5 after learned temporal smoothing; clinician annotations changed at 5.9 ± 1.1.The clinician reference contained 1,774 observed video–sector onsets, with a median bracketing-interval width of 4.3 seconds.Source: §4.6 · p. 8

4.3 Primary Classification Results

• The complete pipeline achieved the highest primary-endpoint accuracy (0.646 ± 0.037), macro-F1 (0.507 ± 0.049), and AUROC (0.882 ± 0.018) among compared methods; versus the emission baseline, accuracy and AUROC means rose 0.046 and 0.021.The paired macro-F1 gain was 0.066 ± 0.022, with improvement in all five folds.Source: §4.3 · p. 8

B. Recall balance, log-loss, and spatial spread

• Median interval error decreased from 15.58 ± 5.03 seconds for emissions to 2.45 ± 1.29 seconds for the full pipeline, approximately 6.4-fold lower.The complete pipeline’s premature-onset proportion was 0.443 ± 0.081.Source: B. Recall balance, log-loss, and… · p. 10

4.7 Class-Specific Recognition

• Full-pipeline F1 improved for several classes, including Scarpa’s fascia (0.229 to 0.455) and bowel (0.217 to 0.312), while preperitoneal fat remained at 0.024.The class plot shows seven of nine classes; Scarpa’s fascia and preperitoneal fat are retained in the table.Source: Figure 3 · p. 10, §4.7 · p. 10, Table 6 · p. 11

4.9 Alternative Spatial Models and Decision Biases

• The hierarchical frame-depth prior with full emission weighting achieved clinician-only accuracy of 0.576 ± 0.039 and macro-F1 of 0.439 ± 0.045.Source: §4.9 · p. 10

4.8 Feature-Context Ablations

• Clinician-only macro-F1 was 0.440 ± 0.045 with compact context, versus 0.402 ± 0.032 without context and 0.428 ± 0.038 with PCA context.For all-label targets, compact context also led, at macro-F1 0.450 ± 0.045.Source: §4.8 · p. 10, Table 7 · p. 12

• The table header lists feature dimensions, accuracy, macro-F1, and macro-AUROC as classification measures.Source: B. All-label classification · p. 12

C. Diagnostic metrics for both label sets

• Spatial stacking reached macro-F1 0.415 ± 0.036, approximately 0.025 below emissions; decision biases scored all-label macro-F1 0.445 versus 0.457 for MRF-only.Source: Table 8 · p. 12, C. Diagnostic metrics for both label… · p. 12

4.10 Parameter Sensitivity, Decoding, and Probability Diagnostics

• MRF macro-F1 improvement persisted across coupling strengths 0.25–1, while tested HMM emission exponents did not improve macro-F1.Viterbi and forward–backward decoding differed by no more than 0.01 on reported label metrics.Source: §4.10 · p. 12, Table 9 · p. 13
• Learned smoothing alone yielded the lowest clinician-only log-loss (1.233 ± 0.157), below emissions (1.350 ± 0.160) and the full pipeline (1.693 ± 0.266).Source: §4.10 · pp. 12–13, Table 9 · p. 13
• Temperature scaling required temperatures of approximately 1.5 for emissions, 11 for the unblended banded HMM, and 70 for the strict HMM, without changing predicted classes, accuracy, or F1.Source: §4.10 · p. 13, Table 9 · p. 13

5.1 Principal Findings

• Accuracy 0.646, macro-F1 0.507, and macro AUROC 0.882 with the complete pipeline, versus 0.600, 0.440, and 0.861 for independent frame classification.Macro-F1 improved in all five folds; flicker decreased from 66.2 to 19.5 changes per sector-minute, and median onset error decreased from 15.58 to 2.45 seconds.Source: §5.1 · p. 13

5.2 Context and Temporal Structure

• Neighbor context improved appearance classification; learned smoothing contributed most AUROC gain and improved Scarpa’s fascia recognition, while spatial regularization moved predictions toward clinician distributions.Source: §5.2 · p. 13
• The strict HMM had the lowest onset error but missed approximately 16% of observed onsets; blending reduced misses to approximately 5%, yet premature onsets remained 44%.The study does not establish whether these improvements are sufficient for clinical use, and smoother predictions may conceal errors at consequential transition points.Source: §5.2 · p. 13

5.3 Methodological Strengths

• Video-grouped evaluation, nested cross-fitting, clinician-drawn primary targets, and inspectable ablations help separate recordings and make contributions and tradeoffs assessable.The compact color descriptors and explicit anatomical constraints provide an interpretable computational baseline for representation-learning studies.Source: §5.3 · p. 13

5.4 Limitations and Future Work

• The retrospective sample was modest, sectors were correlated, and F1 remained 0.024 for preperitoneal fat and 0.312 for bowel.Sparse keyframes limited temporal precision, with a median onset-reference interval of 4.3 seconds; transferred labels were heuristic, and repeated comparisons on the same folds may yield optimistic selection.Source: §5.4 · p. 13
• The complete pipeline had higher log-loss than learned smoothing alone despite superior classification metrics.Source: §5.4 · pp. 13–14
• Future temporal windows and forward–backward inference make the pipeline noncausal, so the results do not establish real-time guidance performance.Prospective evaluation should assess causal prediction, calibration, external generalizability, surgeon decisions, and patient outcomes.Source: §5.4 · p. 14