Cactus: Multidisciplinary Research Overview

• Cactus is a cross-disciplinary term that refers both to acronym-based systems in fields like machine learning and medical imaging, and to specialized mathematical constructs in algebra, geometry, and graph theory.
• Biomedical CACTUS frameworks offer explainable classification by abstracting clinical data into interpretable states, achieving high balanced accuracy and robust feature stability under missingness.
• Additional applications include a high-performance computing environment, precise cardiac ultrasound assessment, advanced timing detectors, and ecological analyses using deep learning.

CACTUS is a term with several distinct technical meanings in contemporary research. In current arXiv literature, it denotes multiple acronymic systems in machine learning, medical imaging, detector instrumentation, and high-performance computing, while the lower-case form “cactus” also names a family of mathematical objects in group theory, algebraic geometry, metric geometry, and graph theory, as well as a subject of agricultural and remote-sensing studies on cactus plants (Andrys-Olek et al., 19 Feb 2026, Elmekki et al., 7 Mar 2025, Degerli et al., 2020, Löffler et al., 2013, Chmutov et al., 2016, Hayamizu et al., 2019, Santos et al., 30 Nov 2025).

1. Principal research usages

The term is best understood as a cross-disciplinary label rather than a single object. In some fields it is an acronym with a fixed expansion; in others it is a structural noun for a restricted kind of graph, group, metric, or scheme. This common naming pattern reflects a recurring emphasis on modularity, constrained combinatorics, or sparse structure, but the underlying formalisms are otherwise unrelated.

Usage	Meaning	Representative reference
CACTUS	Comprehensive Abstraction and Classification Tool for Uncovering Structures	(Andrys-Olek et al., 19 Feb 2026)
CACTUS	Cardiac Assessment and ClassificaTion of UltraSound	(Elmekki et al., 7 Mar 2025)
CACTUS	Cmos ACtive pixel Timing $\mu$Sensor	(Degerli et al., 2020)
Cactus	Open-source HPC framework with “flesh” and “thorns”	(Löffler et al., 2013)
Cactus group	Group generated by interval reversals	(Chmutov et al., 2016)
Cactus metric / cactus graph	Metric or graph realized by a cactus graph	(Hayamizu et al., 2019, Das, 2014)

2. Explainable biomedical machine learning

In biomedical machine learning, CACTUS most prominently denotes Comprehensive Abstraction and Classification Tool for Uncovering Structures, an explainable framework designed for trustworthy decision-making under incomplete clinical data. One formulation targets small, heterogeneous, and incomplete clinical cohorts by combining feature abstraction, interpretable classification, and feature stability analysis under increasing missingness. Continuous variables are abstracted into interpretable Up and Down categories using the ROC curve, classification is performed with a modified Naive Bayes classification algorithm, and robustness is assessed through average relative change in feature importance for the top 10 features and overlap of top 10 features across complete and missing-data variants. In the Haematuria Biomarker cohort, the evaluation used a final cleaned dataset of 568 patients, with 201 bladder cancer (BC) and 367 non-bladder cancer (non-BC) patients, and stress-tested missingness at 10%, 20%, and 30% against Random Forest (RF), AdaBoost, XGBoost (XGB), LightGBM (LGBM), and CatBoost. CACTUS achieved the best overall performance in balanced accuracy and sensitivity across the total, male, and female cohorts, while also showing strong feature stability under degradation of data completeness (Andrys-Olek et al., 19 Feb 2026).

A second biomedical use of the same acronym applies CACTUS to early age-related macular degeneration (AMD) staging. In that setting, CACTUS is an explainable, graph-based classification framework that transforms structured tabular variables into symbolic Up flip and Down flip states, builds class-wise weighted directed graphs of feature flips, and scores class membership with variants denoted CPB, CDG, and CPR. The class score is defined as

$C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$

with prediction by $\arg\max(C)$, and the framework adds a confidence score based on the average absolute separation between the winning class and the others. The AMD study used 29,908 patients and 218 features, deliberately excluding retinal-image-derived features except for the label, and evaluated robustness under 20%, 40%, 60%, and 80% randomly removed values. The CPR (PageRank) variant achieved about 0.34 balanced accuracy on the full dataset and remained around 0.27–0.28 at 80% missingness, outperforming the reported baseline models across all tested missingness levels (Gherardini et al., 16 Jun 2025).

3. Cardiac ultrasound dataset and assessment framework

In echocardiography, CACTUS stands for Cardiac Assessment and ClassificaTion of UltraSound. It was introduced as the first open graded dataset for Cardiac Assessment and ClassificaTion of UltraSound, motivated by the absence of publicly available cardiac ultrasound resources that jointly support view classification and image-quality grading. The dataset was acquired from a CAE Blue Phantom using a GE M4S Matrix Probe and GE Healthcare Vivid-Q ultrasound machine, and contains 37,736 images labeled by cardiac view and by a quality grade from 0 to 10. The views are A4C, SC, PL, PSAV, PSMV, and a Random images class; preprocessing cropped machine overlays, resized images to 224 × 224, and used a 70% training, 10% validation, 20% testing split (Elmekki et al., 7 Mar 2025).

The associated deep learning pipeline is explicitly two-stage. A ResNet18 classifier first predicts the cardiac view using stochastic gradient descent (SGD), categorical cross-entropy (CCE), batch size 128, 30 epochs, and learning rate 0.001. A transfer-learning stage then freezes the shared encoder and adds a grading head trained with Mean squared error (MSE). Reported results include 99.97% training accuracy and 99.43% validation accuracy for classification, grading losses of 0.1154 on training and 0.3067 on validation, and a test grading loss of 0.1077 for the transfer-learning model. In real-time phantom scans, the framework reached classification accuracy up to 96%, and two cardiac imaging experts assigned an average overall score of 8/10 in the questionnaire-based evaluation (Elmekki et al., 7 Mar 2025).

4. Cactus as a high-performance computing framework

In scientific computing, Cactus is an open-source, modular, portable programming environment for collaborative high-performance applications. Its core architectural distinction is between the flesh, which provides compile-time and run-time interfaces, and thorns, which are plug-in modules implementing science or infrastructure functionality. The framework emerged in 1996 at the National Center for Supercomputer Applications (NCSA) and the Albert Einstein Institute, initially to support the numerical relativity community, and later became foundational infrastructure for the Einstein Toolkit. The design separates physics and algorithms from infrastructure concerns such as parallelism, I/O, checkpoint/restart, scheduling, and mesh management; it also uses small domain specific languages (DSLs) for distributed data structures and scheduling, enabling run-time reflection in Fortran and C/C++ (Löffler et al., 2013).

The component model is formalized through the Cactus Configuration Language (CCL). A thorn is specified through three required files—interface.ccl, param.ccl, and schedule.ccl—and two optional files, configuration.ccl and test.ccl. CCL encodes variables, parameters, aliased functions, scheduling, and build-time capabilities, and makes a sharp distinction between interface and implementation, allowing multiple thorns to provide the same interface and to be interchanged at runtime. The paper’s canonical example is the driver interface, implemented by PUGH and Carpet. Standard schedule bins include CCTK_STARTUP, CCTK_PARAMCHECK, CCTK_INITIAL, CCTK_PRESTEP, CCTK_EVOL, CCTK_POSTSTEP, and CCTK_ANALYSIS (Allen et al., 2010).

5. CACTUS as a timing detector concept

In detector instrumentation, CACTUS denotes Cmos ACtive pixel Timing $\mu$Sensor, a radiation-hard depleted monolithic active pixel sensor developed for high-precision charged-particle timing. The device is implemented in a standard $150\,\mathrm{nm}$ high-voltage CMOS process from LFoundry (LF15A) with high-resistivity substrate material, and is motivated by the HL-LHC pile-up regime, where the average number of interactions per bunch crossing is expected to reach about 200. The target operating regime is a timing resolution of the order of tens of picoseconds with spatial granularity around $\sim 1\,\mathrm{mm}$, making CACTUS a monolithic alternative to timing systems based on LGADs (Degerli et al., 2020).

Architecturally, each pixel contains a fast charge-sensitive amplifier (CSA), a leading-edge discriminator, and a 4-bit DAC for threshold compensation. Two pixel variants, $1\times 1\,\mathrm{mm^2}$ and $1\times 0.5\,\mathrm{mm^2}$, were fabricated. Simulations for a thinned $\sim 100\,\mu\mathrm{m}$ sensor with $1.5\,\mathrm{pF}$ detector capacitance and $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$0 total bias current gave a rise time of about $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$1 and an input-referred noise of about $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$2. In the first prototype, measured breakdown voltages were about $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$3 for version A and about $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$4 for version B. Despite an unexpected capacitance increase to more than $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$5, timing measurements with a $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$6 source and PMT reference achieved $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$7 after time-over-threshold correction for a $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$8 thick sensor biased at $C_{c,m} = \sum_{i \in N} \sigma(c, x_i, m),$9 and threshold $\arg\max(C)$0; lower thresholds degraded performance to about $\arg\max(C)$1 at $\arg\max(C)$2 and $\arg\max(C)$3 at $\arg\max(C)$4 (Degerli et al., 2020).

6. Cactus in group theory and algebraic geometry

In algebra and combinatorics, the cactus group is a central object. For the classical cactus group $\arg\max(C)$5, generators are interval reversals $\arg\max(C)$6, $\arg\max(C)$7, subject to

$\arg\max(C)$8

$\arg\max(C)$9

$\mu$0

There is a natural map to the symmetric group, and the pure cactus group is its kernel. This interval-reversal formalism is closely linked to tableau combinatorics and Kazhdan–Lusztig theory: the Berenstein–Kirillov group is shown to be a quotient of the cactus group, and cactus groups admit an alternative presentation in terms of Bender–Knuth generators (Bellingeri et al., 2022, Chmutov et al., 2016).

The same pattern extends far beyond type $\mu$1. For an arbitrary finite Coxeter system $\mu$2, the generalized cactus group $\mu$3 is generated by $\mu$4 indexed by connected spherical subsets, with a natural surjection $\mu$5 and kernel $\mu$6. The generalized pure cactus group embeds into a right-angled Coxeter group, and generalized cactus groups are linear. In the affine setting, the affine cactus group $\mu$7 is described using circular intervals and identified with a generalized cactus group of type $\mu$8; it embeds into a semidirect product $\mu$9, which yields solvability of the word problem, trivial center, absence of odd-order torsion, and torsion-freeness of the pure affine cactus group. On the moduli-theoretic side, the real locus of the compactification $150\,\mathrm{nm}$0 of cactus flower curves has equivariant fundamental group the virtual cactus group, and a degeneration from a twisted real form of $150\,\mathrm{nm}$1 induces a natural homomorphism from the affine cactus group to the virtual cactus group. A further connection sends the cactus group associated with a Coxeter group to the invertible elements of Lusztig’s asymptotic algebra (Yu, 2022, Chemin, 27 Jan 2025, Ilin et al., 2023, Rouquier et al., 2024).

In algebraic geometry and tensor theory, “cactus” describes scheme-theoretic rather than reduced-point decompositions. The cactus rank of a homogeneous polynomial is the smallest length of an apolar zero-dimensional scheme, and algorithmic work extends Hankel/moment methods to recover the support and multiplicities of minimal apolar schemes. The Grassmann cactus variety $150\,\mathrm{nm}$2 generalizes both cactus varieties and Grassmann secant varieties by considering $150\,\mathrm{nm}$3-planes contained in the span of finite subschemes of degree at most $150\,\mathrm{nm}$4; its characterization can be reduced to finite schemes of socle dimension at most $150\,\mathrm{nm}$5. Border apolarity has also been extended from secant varieties to cactus varieties of smooth projective toric varieties, where a border cactus decomposition is a multihomogeneous ideal in the Cox ring that serves as a witness for cactus-variety membership (Bernardi et al., 2018, Buczyńska et al., 29 Jul 2025, Buczyńska et al., 27 Jan 2026).

7. Graph-theoretic, metric, and biological uses

In graph theory, a cactus is a connected graph in which each edge belongs to at most one cycle, or equivalently a connected graph in which every block is either an edge or a cycle. This restricted cycle structure underlies several algorithmic and structural results. A cactus metric is a finite metric realized by an edge-weighted $150\,\mathrm{nm}$6-cactus; every cactus metric has a unique optimal realization, and recognition together with construction of that optimal realization can be performed in $150\,\mathrm{nm}$7 time. On cactus graphs, linear-time algorithms are reported for several classical tasks, including shortest paths, minimum dominating sets, minimum 2-neighbourhood covering sets, labeling problems, and spanning-tree constructions. In infinite-graph theory, minimal edge cuts separating ends can also be encoded by a cactus, extending the classical Dinits–Karzanov–Lomonosov picture for finite graphs (Hayamizu et al., 2019, Das, 2014, Evangelidou et al., 2011).

The biological meaning of “cactus” remains active in agricultural and ecological computation. For prickly pear cultivation, a recent study examines linear Fresnel lenses mounted on an unmanned ground vehicle (UGV) for chemical-free, precision weed-control in tropical semiarid regions. The proposed workflow separates weed mapping during non-optimal solar hours from targeted solar termination during the high-irradiance window, with a practical treatment window of roughly 3.5 hours per day; the method is presented as viable for post-emergent, sparse infestations rather than dense blanket treatment (Santos et al., 30 Nov 2025). In remote sensing, an Enhanced Randomly Initialized Convolutional Neural Network (ERI-CNN) was developed for recognition of the columnar cactus Neobuxbaumia tetetzo in UAV imagery from the Tehuacán-Cuicatlán Valley. Using 21,500 images from the “Cactus Aerial Photos” dataset, the study reported 98% test accuracy, 97% precision, 97% recall, 97.5% F1-score, and 0.056 loss (Atitallah et al., 2021).