A new study published in BMC Medical Imaging by Suxiang Yu, Xiaoqin Yang, Yihe Duan, Dun Hua, Bai Yun, Huimiao Sun, Lingling Zhang, Feihong Wu, Dandan Yang, and Xin Huang takes aim at exactly this problem. The team, spanning pathology departments and engineering schools in China, evaluated and adapted Cellpose-SAM, a modern generalist cell segmentation model, to the specific and stubborn challenge of instance segmentation in overlapping cervical cytology images. Their results, while modest in absolute terms, point to a practical lesson for the field: even powerful foundation models need careful, task-specific tuning before they can be trusted on the messy realities of clinical slides.

Cellpose-SAM itself represents an interesting convergence of two research threads. Cellpose, originally released in 2020, popularized a two-step approach to cell segmentation in which a neural network first predicts the flow of gradients from every pixel toward the center of its parent cell, and an algorithm then follows those flows to group pixels into individual instances. This design elegantly sidesteps the need for the network to label each cell with a unique identifier, which is notoriously difficult to train. The SAM in the name refers to the Segment Anything Model, a vision foundation model trained by Meta AI on an enormous corpus of natural images. By marrying SAM’s pretrained visual backbone to Cellpose’s flow-based decoding, Cellpose-SAM inherits strong generalization to cell types and imaging modalities it has never seen before, including microscopy styles far removed from its training distribution.

But strong generalization is not the same as optimal performance on any given task. As the authors note in the study, the default configuration of Cellpose-SAM may not be suited to densely overlapping cervical cytology images, where cell adhesion, weak cytoplasmic contrast, and ambiguous boundaries conspire to obscure individual cell structures. When two overlapping epithelial cells stain nearly identically and share a barely perceptible junction, the model must decide where one cell ends and the other begins, a judgment that even the underlying flow field may represent poorly without further training on comparable data.

To give the model that training, the researchers assembled a single-center retrospective dataset from the Fourth Central Hospital of Baoding in Hebei Province, China. The dataset consisted of 433 images of overlapping cervical cells, each accompanied by expert manual annotations that served as the ground truth against which the algorithm’s outputs were measured. These images were divided into training, validation, and independent test sets at a ratio of 70 percent, 10 percent, and 20 percent, respectively. The split is a standard but important design choice: the training set teaches the model, the validation set guides tuning decisions, and the test set, held completely apart from both, provides an honest estimate of how the adapted system performs on data it has never encountered during development.

The adaptation strategy had two complementary components. The first was fine-tuning, in which the pretrained Cellpose-SAM weights were updated on the cervical cytology training images, allowing the model’s internal representations to adjust to the specific staining patterns, cell morphologies, and clutter of Pap-smear preparations. The second was inference-parameter optimization, a step that is often overlooked but can be as consequential as retraining itself. Cellpose-SAM exposes several knobs at prediction time: the expected cell diameter, which calibrates the scale at which the network looks for objects; the flow threshold, which determines how strictly the gradient-following algorithm trusts its flow predictions when assembling cell instances; and the cell probability threshold, which decides how readily the model declares a pixel to belong to a cell at all. Rather than accepting defaults, the team systematically tuned these parameters on the validation set, using it as a proxy for the true target distribution and then freezing the chosen values for evaluation.

Performance was assessed on the independent test set using four established segmentation metrics. The Dice coefficient, which measures the harmonic mean of overlap between predicted and annotated regions, reached 0.8737 for the adapted model. Intersection over Union, a stricter spatial-overlap measure, came in at 0.7199. The Aggregated Jaccard Index, a metric designed specifically for instance-level segmentation because it penalizes both missed and spurious instances across a whole image, registered 0.7024. The error rate, which captures instances the model failed to segment correctly, was 0.1551. The authors performed statistical comparisons between the default and adapted configurations using paired image-level analysis, and the adapted model showed improved performance across these measures, suggesting better agreement with the expert manual annotations in the overlapping-cell regions.

The numbers deserve careful interpretation. A Dice coefficient of roughly 0.87 indicates substantial pixel-level agreement, and an IoU above 0.72 is respectable for a task where the ground truth itself is contested by virtue of ambiguous cell boundaries. The AJI of about 0.70, meanwhile, is the more sobering figure, because it aggregates performance at the level of whole instances: every cell the model misses entirely or invents from nothing drags the score down across the entire image. An error rate of about 15.5 percent means that roughly one in six or seven predicted instances does not match the annotation well. In other words, the adapted Cellpose-SAM is a capable assistant rather than a finished autonomous cytologist, and the gap between 0.87 pixel overlap and 0.70 instance accuracy quantifies precisely the residual difficulty of separating cells that touch and overlap.

What makes the study noteworthy methodologically is its insistence that off-the-shelf foundation models are starting points, not endpoints. The cervical cytology community has seen a wave of deep-learning screening tools, many of which treat segmentation as a solved problem and focus instead on classification of abnormalities. Yet if the segmentation stage silently merges overlapping cells or slices one cell into two, every downstream measurement, from nucleus-to-cytoplasm ratio to nuclear area, inherits that distortion. The Baoding team’s finding that simply tuning the cell diameter, flow threshold, and cell probability threshold on a small validation set yielded measurable gains over defaults suggests that laboratories adopting Cellpose-SAM or similar tools should budget time and annotation effort for exactly this kind of calibration, tailored to their staining protocols and scanner optics.

The authors are appropriately measured about the limits of their work. The dataset comprised 433 images from a single center, which raises questions about how well the adapted model would transfer to slides prepared with different stains, scanned on different instruments, or drawn from patient populations with different distributions of inflammation, atrophy, and glandular cells. They state explicitly that further external validation and broader benchmarking are required before clinical generalization can be established. The retrospective study was approved by the hospital’s ethics committee under approval number 2022011, conducted in accordance with the Declaration of Helsinki, and used de-identified, anonymized images, with the informed-consent requirement waived for the retrospective design. The work was funded by the S&T Program of Hebei under project number 22377774D, and the authors declared no competing interests.

Still, the trajectory is easy to read. Foundation models pretrained on millions of natural images are rapidly becoming the default starting point for biomedical image analysis, and the interesting research questions are shifting from how to build such models to how to adapt them efficiently to narrow, high-stakes domains. Cervical screening is a particularly attractive target: it is a globally distributed task with a persistent shortage of trained cytologists, and the cost of a missed abnormality is measured in human lives. If modest fine-tuning and parameter tuning can push a generalist segmentation model to useful accuracy on the notoriously tangled cells of a Pap smear, the path from laboratory benchmark to screening-room assistance becomes considerably more concrete, and the 433 clumped, ambiguous, stubbornly overlapping cells of Baoding become a small but meaningful waypoint on that road.

Subject of Research: Instance segmentation of overlapping cervical cells in Pap-smear images using adapted Cellpose-SAM

Article Title: Adapting Cellpose-SAM for instance segmentation of overlapping cervical cells in Pap-smear images

Article References: Yu, S., Yang, X., Duan, Y., Hua, D., Yun, B., Sun, H., Zhang, L., Wu, F., Yang, D., & Huang, X. (2026). Adapting Cellpose-SAM for instance segmentation of overlapping cervical cells in Pap-smear images. BMC Medical Imaging. https://doi.org/10.1186/s12880-026-02845-8

Image Credits: AI Generated

DOI: 10.1186/s12880-026-02845-8

Keywords: Cellpose-SAM, cervical cytology, Pap smear, instance segmentation, deep learning, fine-tuning, medical imaging, computer vision, cervical cancer screening, parameter adaptation, BMC Medical Imaging, Adapting

Robotic-assisted surgery changes the physical and visual demands placed on surgeons. Instead of manipulating instruments directly through small abdominal incisions, the surgeon controls articulated tools from a console, typically while viewing a magnified three-dimensional image. The system can provide greater instrument articulation and fine motion control, but these advantages do not automatically translate into technical competence. Suturing requires coordinated needle handling, accurate tissue bites, controlled tension, and efficient instrument exchanges. A trainee must also work within the constraints of a narrow operative field while maintaining a stable camera view and avoiding unnecessary movements. Laparoscopic experience provides a foundation, yet robotic instruments and the console interface create a different motor environment. The authors therefore treated adaptation as more than simply repeating laparoscopic exercises with a robot. They redesigned the existing tasks for robotic instrumentation and added a non-dominant forehand suturing exercise intended to challenge a less familiar hand position.

The original ATLAS curriculum was developed to train and assess advanced laparoscopic suturing through defined tasks and proficiency standards. Its central idea is mastery learning: learners practice until they meet an objective performance level, rather than stopping after a fixed number of attempts or a predetermined amount of time. This approach can make training more consistent because progress is linked to demonstrated ability. It also allows educators to identify specific technical weaknesses and provide targeted practice. For R-ATLAS, the investigators adapted all six ATLAS tasks to the Intuitive Abdominal Dome Trainer, a simulator designed to reproduce aspects of abdominal surgery. The additional task, called 3ND, required non-dominant forehand suturing. Such a task is technically important because surgeons may need to use either hand depending on anatomy, instrument position, access angle, or the direction of a repair. The resulting curriculum contained seven exercises intended to represent demanding components of robotic suturing.

To establish reference performance, four expert robotic surgeons completed five repetitions of every task using a da Vinci Xi system. Their performances were recorded on video and scored independently. In total, the study collected 140 attempts. Six attempts were classified as outliers because they fell more than two standard deviations from the relevant performance distribution. After those exclusions, 134 attempts remained for analysis. The investigators used descriptive statistics to summarize how the experts performed and to calculate benchmarks for each exercise. This design does not compare novices with experts, and it does not test whether achieving a benchmark improves outcomes in the operating room. Instead, it uses expert performance as a starting point for defining what a high-level simulated performance might look like. The process reflects a common strategy in skills education: first establish a measurable reference range, then examine whether the standard is reliable, teachable, and related to real clinical performance.

Benchmarks were established for all seven R-ATLAS tasks, but the results also exposed substantial differences among the experts. Mean performance varied for Tasks 1 through 5 and for the 3ND exercise. Task 6, by contrast, showed relatively similar performance across the surgeons. This pattern challenges the idea that expertise always produces one narrow, uniform technique. Surgeons can reach a technically acceptable result through different combinations of movement, timing, instrument positioning, and needle control. Variation may also reflect the intrinsic difficulty of a task or the ways in which specialists develop individual strategies over years of practice. For educators, the finding is both useful and cautionary. A benchmark derived from a small expert group can provide a practical target, but it may also encode the particular mix of styles represented in that group. The researchers therefore describe the thresholds as preliminary and call attention to the need for further validation before they are treated as definitive standards.

The study also examined perceived workload using the NASA Task Load Index, a tool that captures subjective demands such as mental effort, physical effort, time pressure, frustration, and perceived performance. Scores ranged from 15.5 to 30 across the exercises. The highest workload was reported for Task 3ND, the non-dominant forehand exercise, while Task 4 produced the lowest workload. These measurements add an important dimension to technical scoring. Two tasks may appear similar when judged by completion time or errors, yet require very different levels of concentration and effort. A high-workload exercise may reveal where trainees are most likely to struggle, even if they eventually complete it successfully. It could also help instructors sequence a curriculum, introduce deliberate practice, or monitor whether repeated training makes a task feel less demanding. Because the workload findings came from the expert sessions, they should not be assumed to represent novice experience, but they identify areas that deserve attention in later studies.

R-ATLAS could eventually support more standardized robotic education by giving programs a shared vocabulary for advanced suturing. A trainee’s progress could be tracked across repeated attempts, with feedback tied to observable performance rather than general impressions. Video recording also creates opportunities for independent review and remote assessment, potentially allowing instructors at different institutions to examine the same technical behaviors. The framework may be especially relevant as surgical training programs integrate robotic procedures while still needing to teach fundamental principles of tissue handling and repair. However, a simulator cannot reproduce every feature of an operation. Real patients introduce variable anatomy, tissue fragility, bleeding, unexpected findings, team communication, and time-sensitive decisions. Meeting a simulated benchmark should therefore be viewed as evidence of performance on a defined exercise, not proof that a surgeon is ready to perform an entire procedure independently.

The authors’ most important message may be that robotic proficiency requires measurement without oversimplification. R-ATLAS supplies an organized platform, seven tasks, expert-derived reference scores, and workload data, but it does not end the debate over what mastery means. Future research will need to test the curriculum with larger and more diverse expert groups, determine how consistently different evaluators score performance, and examine how trainees improve with practice. Studies could also investigate whether benchmark achievement transfers to clinical skills, whether different robotic platforms produce comparable results, and how patient-specific complexity should influence assessment. For now, the study offers a practical bridge between laparoscopic education and robotic surgery. By translating an advanced suturing curriculum into a robotic environment while acknowledging expert variability, R-ATLAS provides educators with a measurable starting point for training surgeons to make precise, controlled movements when the smallest technical details can matter most.

An important feature of the R-ATLAS design is its attempt to preserve the educational logic of the original ATLAS curriculum while changing the interface through which the skills are performed. This distinction is relevant to curriculum design: a robotic simulator can assess suturing mechanics, but the meaning of a score depends on the task’s construction, the platform, and the scoring rules. By adapting the exercises specifically for robotic instruments, the investigators created a platform-focused assessment rather than assuming that laparoscopic standards could be transferred unchanged. The work therefore addresses a practical gap identified in robotic education, where programs have adopted robotic technology faster than they have developed consistent approaches for teaching advanced technical maneuvers.

The benchmark process also illustrates why proficiency standards require ongoing validation. Four experts completing repeated trials can reveal the range of performance expected from highly experienced users, but that sample is not large enough to establish how broadly the thresholds apply across specialties, institutions, training backgrounds, or robotic systems. Removing six predefined outlier attempts reduces the influence of unusually atypical performances, yet it can also narrow the observed distribution if those attempts reflect meaningful variation rather than measurement noise. The resulting scores should consequently be interpreted as provisional estimates derived from this study’s expert sample. Reliability testing, including agreement among independent raters and consistency across assessment sessions, would strengthen the evidence that a trainee’s score reflects skill rather than scoring or testing variability.

The findings are also consistent with a broader principle in simulation-based mastery learning: assessment is most useful when it is connected to deliberate practice and actionable feedback. A numerical threshold can tell an instructor that performance falls short, but it does not by itself identify whether the problem involves needle orientation, tissue handling, economy of motion, or control of the non-dominant instrument. R-ATLAS may become more educationally valuable if future implementations pair its benchmarks with error taxonomies, motion-based measures, or structured video feedback. Such additions could help distinguish a slow but precise learner from a fast performer whose technique creates unnecessary force or inconsistent suture placement.

Clinical transfer remains the central question for any simulator-based benchmark. Prior simulation research cited by the investigators supports the general proposition that proficiency-based training can improve technical performance, but the present study does not demonstrate that R-ATLAS scores predict patient outcomes or operating-room readiness. Establishing that relationship would require prospective studies following trainees from simulator practice into clinical cases and examining outcomes with appropriate safeguards. It would also be important to determine whether repeated practice produces durable retention rather than short-term familiarity with the simulator. Even before those studies are completed, the curriculum offers a research-ready structure for comparing training strategies and for studying how robotic dexterity develops, making its preliminary benchmarks useful as measurement tools as well as educational targets.

Subject of Research: Robotic-assisted suturing training and proficiency benchmarking

Article Title: Robotic ATLAS: adapting advanced laparoscopic suturing training to a robotic platform with proficiency benchmark scores

Article References: Jonas, N., Chen-Goodspeed, A., Yousef, S., Hsu, C.-H., Soliman, D., Nepomnayshy, D., Zheng, J., Ford, H., Nejad, A., Ritter, M., Hodges, J., & Ghaderi, I. (2026). Robotic ATLAS: adapting advanced laparoscopic suturing training to a robotic platform with proficiency benchmark scores. Global Surgical Education – Journal of the Association for Surgical Education, 5(1), Article 171. https://doi.org/10.1007/s44186-026-00572-w

Image Credits: AI Generated

DOI: 10.1007/s44186-026-00572-w

Keywords: Robotic surgery, Surgical education, Laparoscopic suturing, Simulation training, Proficiency benchmarks, R-ATLAS, Surgical robotics, NASA-TLX, Robotic, ATLAS, adapting, advanced