RSDNet: Intraoperative Surgery Duration Estimation

• RSDNet is a CNN-LSTM architecture that estimates remaining surgery duration from laparoscopic video using automatically generated progress labels.
• It employs a two-stage training strategy with progress and RSD regression, leveraging temporal cues without relying on manual phase annotations.
• Empirical results on Cholec120 and Bypass170 demonstrate superior performance over naive baselines and competitive accuracy compared to phase-based approaches.

RSDNet is an overloaded acronym in the arXiv literature rather than a single canonical architecture. In the earliest and most explicit usage in the provided corpus, it denotes a CNN-LSTM pipeline for intraoperative remaining surgery duration estimation from laparoscopic video without manual annotations (Twinanda et al., 2018). Closely related spellings later identify a BYOL-based remote-sensing representation learner and a minimum-divergence neural classifier grounded in the $S$-divergence family (Jain et al., 2022, Jana et al., 18 Mar 2026). The name is also regularly confused with distinct abbreviations such as RDSNet and RFDN, which address different problems and are not equivalent to RSDNet (Wang et al., 2019, Liu et al., 2020).

1. Terminological scope and disambiguation

The acronym appears in several unrelated subfields. For precise reading, the arXiv context matters more than the string match.

Name	Domain	Defining note
RSDNet	Surgical video analysis	Remaining surgery duration from laparoscopic video without manual annotations (Twinanda et al., 2018)
RSDnet / RS-BYOL	Remote sensing SSL	BYOL-style invariant representation learning from multi-spectral and SAR images (Jain et al., 2022)
rSDNet	Robust classification	Minimum-divergence neural learning against label noise and adversarial attacks (Jana et al., 18 Mar 2026)
RDSNet	Detection and segmentation	Reciprocal object detection and instance segmentation, not RSDNet (Wang et al., 2019)
RFDN	Image super-resolution	Residual Feature Distillation Network, explicitly not “RSDNet” (Liu et al., 2020)
RSD	Diffusion SR	Residual Shifting Distillation, not RSDNet (Selikhanovych et al., 17 Mar 2025)

This multiplicity is not merely lexical. Each variant targets a different statistical problem, uses a different supervisory signal, and adopts a different architectural prior. Consequently, citations that mention “RSDNet” without an arXiv identifier are potentially ambiguous.

2. Clinical problem setting of the original RSDNet

In (Twinanda et al., 2018), RSDNet is introduced for intraoperative prediction of remaining surgery duration during laparoscopic procedures. The problem is motivated by operating-room planning: accurate duration estimates improve anesthesia planning, patient safety, and operating-room utilization, whereas inaccurate estimates lead to delays, overtime, and wasted resources. The paper therefore shifts attention from preoperative prediction toward continuous intraoperative estimation from the live visual stream produced by the laparoscope.

The task is defined as follows: given a laparoscopic video up to the current time, estimate how many minutes remain until the surgery ends. The paper emphasizes that this is more actionable than a single upfront estimate because it supports continuous updates during surgery. The central difficulty is that identical remaining-duration values may correspond to very different visual scenes across cases owing to variability in patient condition, surgeon style, workflow pace, and intraoperative events. RSDNet is based on the hypothesis that laparoscopic frames nevertheless contain enough discriminative visual cues for a deep model to infer surgical progress and, by extension, remaining time (Twinanda et al., 2018).

A defining property of the method is that it does not depend on manual phase annotations. Earlier pipelines relied on manually labeled phases or tool annotations, which are expensive and difficult to scale across procedure types. RSDNet instead uses labels that can be generated automatically from timestamps: progress labels normalized from $0$ at the beginning of surgery to $1$ at the end, and remaining-surgery-duration labels measured in minutes.

3. Architecture, supervision, and optimization

The original RSDNet adopts a two-stage CNN + LSTM pipeline. Laparoscopic videos are sampled at $1$ fps. A ResNet-152 backbone performs frame-level feature extraction, and an LSTM models temporal evolution over the extracted sequence. The network is trained for multi-task prediction of both progress and remaining surgery duration (RSD), with the LSTM output concatenated with the elapsed time before the final RSD regression layers (Twinanda et al., 2018).

The training strategy is deliberately staged because end-to-end optimization of the full ResNet + LSTM model is memory-intensive. First, the CNN is finetuned to regress progress from individual frames. This encourages the visual backbone to learn semantically meaningful surgical features without requiring manual phase labels. Second, CNN features are frozen or extracted and the LSTM is trained on full sequences. Among the four CNN task formulations examined—RSD classification, progress classification, RSD regression, and progress regression—the paper reports that progress regression works best. The rationale given is that progress is normalized across videos, making frames from different surgeries more comparable, whereas RSD values are visually less consistent across cases (Twinanda et al., 2018).

The formulation is explicitly multi-task. The paper states the temporal relation

$t_{rsd} = T - t_{el},$

where $t_{rsd}$ is remaining surgery duration, $T$ is total surgery duration, and $t_{el}$ is elapsed time. Both progress and RSD are predicted directly, rather than deriving RSD solely from progress. The losses for both regression heads are smooth $L_1$ losses with equal weighting: $$\mathcal{L} = \mathcal{L}_{prog} + \mathcal{L}_{rsd}.$$ RSD targets are normalized during training to stabilize regression, with $0$0 for cholecystectomy and $0$1 for bypass, and predictions are denormalized back to minutes at test time (Twinanda et al., 2018).

This design has two notable implications. First, progress serves as a surrogate supervision signal that is automatically available from timestamps. Second, the direct RSD head allows the model to exploit temporal and visual regularities that are not captured by a deterministic transformation of predicted progress alone.

4. Datasets, baselines, and empirical behavior

The principal evaluation in (Twinanda et al., 2018) spans two datasets with markedly different temporal structure. Cholec120 contains 120 cholecystectomy videos from 33 surgeons, combines Cholec80 with 40 additional videos, totals about 75 hours, and has average duration 38.1 ± 16.0 min. Videos were recorded at 25 fps, evaluated with 4-fold cross validation, and processed at 1 fps for training and testing. Bypass170 contains 170 gastric bypass videos, over 327 hours, 28 surgeons, and average duration 115 ± 29 min. For bypass, the same T1/T2/V/E split style is used, and the paper varies CNN training data size among 40, 60, and 80 videos while using all 120 training videos for LSTM training (Twinanda et al., 2018).

The paper compares RSDNet against a naive duration prior, phase-inferred approaches, TimeLSTM, a single-task LSTM, and a progress-derived RSDNet baseline. On Cholec120, performance is reported as mean absolute error in minutes:

• Naive mean: 11.1 ± 8.0
• Naive median: 10.7 ± 8.0
• Phase-inferred GT: 8.1 ± 5.8
• Phase-inferred LSTM: 8.5 ± 6.2
• TimeLSTM: 7.7 ± 5.2
• Single-task: 8.5 ± 5.2
• RSDNet (progress-derived): 9.6 ± 4.7
• RSDNet: 8.1 ± 5.4 (Twinanda et al., 2018)

On Bypass170, only the naive baselines and RSDNet are compared because phase labels are unavailable:

• Naive mean: 22.5 ± 15.6
• Naive median: 21.3 ± 16.5
• RSDNet (40 CNN / 120 LSTM): 15.6 ± 7.9
• RSDNet (60 / 120): 15.8 ± 7.5
• RSDNet (80 / 120): 17.0 ± 8.1 (Twinanda et al., 2018)

These results support three claims made in the paper. RSDNet is significantly better than the naive baseline without using manual annotation; direct RSD prediction is better than deriving RSD from progress; and, on cholecystectomy, RSDNet is comparable to phase-based methods while avoiding manual annotations. The bypass results further suggest that the same weakly supervised template transfers to a longer and more variable procedure class.

5. Generalization, ablations, and interpretability

A major contribution of the surgical RSDNet paper is its argument for scalability across surgery types. Because training labels come from timestamps rather than expert phase annotation, the method is presented as more easily transferable to new procedures. The bypass study is the concrete test case: the approach moves from cholecystectomy to gastric bypass, which is substantially longer and more variable, while still outperforming naive duration priors by a wide margin (Twinanda et al., 2018).

The ablation findings are also specific. Using only ImageNet-pretrained ResNet features without finetuning degrades performance substantially: on Cholec120, the MAE becomes 12.3 ± 6.4, which is worse than the full RSDNet and even worse than the naive method in some settings. The multi-task LSTM that jointly predicts progress and RSD performs slightly better than the single-task LSTM that predicts only RSD. Direct RSD regression outperforms the RSDNet (Progress-Derived) baseline. On Bypass170, training the CNN with 40 or 60 videos works better than using 80 videos in that setup, which the authors interpret as evidence that the two-step optimization and CNN finetuning size matter (Twinanda et al., 2018).

The paper also devotes unusual attention to interpretability. Examination of LSTM hidden-cell activations reveals several classes of learned temporal cues: monotonic progress cells that increase or decrease over time; midpoint-sensitive cells that appear to split videos around roughly $0$2 completion; and a “specimen bag detector,” where strong activation accompanies the appearance of the specimen bag and predicted RSD drops. The authors present these visualizations as evidence that the model is not simply memorizing duration priors, but is learning semantically meaningful cues associated with workflow progression and end-of-case events (Twinanda et al., 2018).

A plausible implication is that the method’s strongest utility lies in regimes where duration priors are least informative. The paper explicitly notes that short and long surgeries are where visual modeling helps most, whereas medium-duration procedures reduce the gap among methods.

6. Other uses of the name: remote sensing and robust classification

A distinct use of the name appears in remote sensing. In (Jain et al., 2022), RSDnet refers to a remote-sensing distillation network based on BYOL, also named RS-BYOL, for self-supervised learning from multi-spectral (MS) and synthetic aperture radar (SAR) imagery. The two views are not RGB augmentations but modality-paired observations of the same scene: MS and SAR act as implicit augmentations because they preserve land-cover semantics while differing in sensing mechanism, spectral response, noise characteristics, and resolution. The architecture follows BYOL with a ResNet50 encoder, 2-layer MLP projection and prediction heads of hidden size 256, and an EMA teacher update

$0$3

with $0$4. The model is pretrained on Sen12MS, using about 90K paired samples from 50% of the dataset in the base setup, with 400 epochs, SGD, cosine warm restart, and batch size 32. In linear evaluation, strong single-channel features reach about 0.92 F1 on EuroSAT, and some bands reach up to 59.6 mIoU on the DFC segmentation task. The paper further reports that MS + SAR pretraining outperforms MS-only pretraining and that RS-pretrained SSL weights outperform ImageNet-supervised weights on many non-RGB and SAR settings (Jain et al., 2022).

An entirely different lowercase variant, rSDNet, is introduced in (Jana et al., 18 Mar 2026) for robust neural classification under label noise and adversarial attacks. This method replaces categorical cross-entropy with a minimum-divergence objective from the $0$5-divergence family. For a $0$6-class classifier with softmax outputs $0$7, the training loss is parameterized by $0$8 and $0$9, with $1$0 controlling down-weighting of outliers and $1$1 refining the robustness-efficiency tradeoff. The paper establishes Fisher consistency, classification calibration implying Bayes optimality, a bounded excess-risk guarantee under uniform label noise, and influence-function-based local robustness to infinitesimal feature contamination. Experiments on MNIST, Fashion-MNIST, and CIFAR-10 show competitive clean-data accuracy and improved robustness, with a recommended default region of $1$2 and $1$3 (Jana et al., 18 Mar 2026).

These two later usages share only the acronymic surface form. Their objectives, data regimes, and theoretical commitments differ completely from the surgical-video RSDNet.

7. Adjacent acronyms and recurrent confusion

Misidentification with nearby acronyms is common. The 2020 super-resolution paper (Liu et al., 2020) is often mistyped in this way, but the model there is explicitly Residual Feature Distillation Network (RFDN), with an enhanced variant E-RFDN; the paper states that it is not “RSDNet”. Its contributions are feature distillation connections (FDC) and shallow residual blocks (SRB) for lightweight single-image super-resolution, and it reports winning first place in the AIM 2020 efficient super-resolution challenge (Liu et al., 2020).

Similarly, RDSNet in (Wang et al., 2019) stands for Reciprocal Object Detection and Instance Segmentation. It uses a two-stream architecture with an object stream and a pixel stream, alternating fusion, a correlation module, a cropping module, and a Mask-Based Boundary Refinement Module (MBRM). This is a detection-and-segmentation architecture rather than a surgery-duration model, even though the spelling differs by only one letter (Wang et al., 2019).

Other nearby abbreviations are also distinct. RDN denotes Residual Dense Network for image restoration (Zhang et al., 2018). RDRNet denotes Reparameterizable Dual-Resolution Network for real-time semantic segmentation (Yang et al., 2024). RSD in (Selikhanovych et al., 17 Mar 2025) abbreviates Residual Shifting Distillation for one-step diffusion-based image super-resolution, not RSDNet. For encyclopedia usage, the arXiv identifier is therefore the decisive disambiguator.

Taken together, the literature supports a narrow and a broad reading of the term. In the narrow sense, RSDNet is the 2018 CNN-LSTM framework for predicting remaining surgery duration from laparoscopic video without manual annotations (Twinanda et al., 2018). In the broad sense, it is an acronym family that has been repurposed for remote-sensing self-supervision and robust neural classification, while remaining adjacent to several non-equivalent abbreviations in vision and image restoration (Jain et al., 2022, Jana et al., 18 Mar 2026).