SCKansformer: Domain-Specific Transformer Models

• SCKansformer is a suite of specialized transformer models designed for high-precision tasks in medical imaging, plasma dynamics, and superconducting systems.
• It integrates innovative techniques such as edge-specific Kolmogorov-Arnold Networks, hierarchical feature pruning, and hybrid global-local attention to enhance model accuracy and interpretability.
• Validated across diverse datasets, SCKansformer consistently delivers state-of-the-art performance in bone marrow cell classification, tokamak skin effect modeling, and high-current superconducting transformer operations.

SCKansformer is a term denoting several distinct, domain-specific systems and models, each characterized by specialized transformer or transformation architectures targeting highly technical applications. The principal contemporary usages of SCKansformer encompass three areas: (1) fine-grained bone marrow cell microimage classification, (2) lumped-parameter modeling of tokamak transformer dynamics including skin effect, and (3) superconducting transformer systems for high-current laboratory operations. This article develops these three meanings in technical depth, referencing the underlying mathematical formulations and validation protocols.

1. Fine-Grained Classification Model for Bone Marrow Cells

SCKansformer designates a novel transformer-based neural architecture optimized for cytological microimage classification, as presented in "SCKansformer: Fine-Grained Classification of Bone Marrow Cells via Kansformer Backbone and Hierarchical Attention Mechanisms" (Chen et al., 2024). The motivation arises from clinical demand for accurate and efficient discrimination among nearly 40 morphologically subtle white-blood-cell subtypes in bone marrow smears.

Standard transformer models, particularly Vision Transformers (ViTs), exhibit several deficiencies in this context: under-expressive nonlinear mappings due to fixed activation functions, large parameter and memory footprints, and poor interpretability of learned nonlinear relationships. SCKansformer addresses these by integrating:

• A Kansformer Encoder replacing the MLP block with a Kolmogorov-Arnold Network (KAN), in which each connection employs a trainable, edge-specific univariate activation $\varphi_{q,p}(\cdot)$.
• An SCConv Encoder containing a Spatial Reconstruction Unit (SRU) to prune spatial redundancies via weighted channel gating, and a Channel Reconstruction Unit (CRU) that combines feature branches adaptively through softmax-weighted fusion.
• A Global-Local Attention Encoder combining Multi-Head Self-Attention (MSA) for global dependencies with a Local Part module—employing shallow depthwise-separable convolutions and h-swish activations—to capture fine-grained local patterns critical for cell morphology.

2. Kansformer Encoder: Kolmogorov-Arnold Network Substitution

The Kansformer Encoder replaces the fixed nonlinearity and linear weights of a standard MLP with a learnable family of univariate functions over each input-output connection:

$$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$

where $\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$ is parameterized and updated during training. This scheme enables:

• Configuration of highly nontrivial, edge-specific nonlinear transformations.
• Parameter efficiency, as fewer hidden units are needed to reach equivalent expressiveness.
• Post-hoc interpretability—each $\varphi_{q,p}$ may be analytically examined (e.g., via Taylor expansion) to elucidate the influence structure between features.

In the architecture, the standard transformer block output

$X_{\text{out}} = X' + \text{FFN}(\text{LN}(X'))$

is replaced by

$X_{\text{out}} = X' + \text{KAN}(\text{LN}(X'))$

where $\text{KAN}$ applies the learned set of scalar functions.

3. SCConv Encoder: Hierarchical Feature Redundancy Reduction

The SCConv Encoder operates on tensors $X\in\mathbb{R}^{N\times C\times H\times W}$ to address spatial and channel feature redundancy:

• Spatial Reconstruction Unit (SRU): Computes group-normalized channel weights $\gamma$ and applies threshold gating to segment features into informative and redundant submaps, followed by cross-reconstruction merging.
• Channel Reconstruction Unit (CRU): Splits channels into $X_{\text{up}}$ (rich branch) and $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$0 (complement branch), processes them through grouped pointwise and global convolutional blocks, and fuses their outputs with adaptive weighting $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$1 computed via global average pooling.

This enforces focus on discriminative spatial regions and channels, reducing overfitting and optimizing data utilization.

4. Performance Evaluation, Datasets, and Comparative Results

SCKansformer was validated on the BMCD-FGCD dataset (92,335 images, 39 classes), the ALL-IDB (260 images), and PBC (17,092 images) datasets. The experimental protocol included ImageNet-1K pretraining, image size standardization to $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$2, and use of the Adam optimizer with cosine annealing. Metrics demonstrate that SCKansformer establishes state-of-the-art performance:

Dataset	Method	Precision	Recall	F1	Accuracy
BMCD-FGCD	SCKansformer	85.82	84.28	84.34	83.23
	WBC-GLAformer	79.82	79.38	79.60	79.45
	ViT	77.65	77.41	77.53	77.42
PBC	SCKansformer	99.47	99.46	99.46	99.46
	WBC-GLAformer	99.40	99.39	99.39	99.40
ALL-IDB	SCKansformer	99.90	99.85	99.85	99.86
	WBC-GLAformer	99.15	99.10	99.09	99.10

The model yields superior gains over all baselines, with +5.52 pp in precision and +3.78 pp in accuracy over the best-performing alternative (WBC-GLAformer) on BMCD-FGCD (Chen et al., 2024).

5. SCKansformer: Tokamak Skin Effect Transformer Model

In plasma control, SCKansformer also refers to a lumped-parameter model of the tokamak transformer that incorporates the slow magnetic-flux penetration (skin effect) in the plasma circuit (Romero et al., 2012). The formulation is based on exact conservation laws:

• Total plasma inductance: $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$3 (internal $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$4, external $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$5)
• Coupled ODEs for current $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$6 and inductance $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$7, including the effect of non-inductive current drive $$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$8 and loop voltages at plasma boundary ($$[Z_{k+1}]_q = \sum_{p=1}^{n_k} \varphi_{k,q,p}([Z_k]_p)$$9), resistive surface ($\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$0), and equilibrium surface ($\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$1):

$\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$2

$\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$3

• Closure for $\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$4 is achieved by a three-point discretization of the loop-voltage profile and a second-order lumped ODE, with parameters empirically determined via system identification on Random Binary Signal (RBS) plasma modulation experiments in the TCV tokamak.

The model demonstrates approximately 70% fit in plasma current prediction under Ohmic conditions, without requiring full 1D transport equations, and is amenable to control-affine reduction for feedback design (Romero et al., 2012).

6. SCKansformer in Superconducting Transformer Systems

A further usage of SCKansformer as "SCT" addresses high-current, low-loss superconducting transformers for laboratory cable/magnet testing. Key technical parameters include:

• Primary: multifilamentary NbTi, $\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$5 H.
• Secondary: NbTi Rutherford cable (parallel configuration), $\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$6 H.
• High current ratio: $\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$7.
• Maximum output: $\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$8 kA DC at $\varphi_{k,q,p}:\mathbb{R}\to\mathbb{R}$9 K.
• Minimal secondary joint resistance: $\varphi_{q,p}$0.
• Quench detection (threshold $\varphi_{q,p}$1 A@200 ms), robust heater-triggered protection, and redundant current diagnostics (Rogowski coil and Hall probe), as validated in practical operation (Yu et al., 2023).

These systems ensure safe, stable current supply for critical superconducting cable assessment, with emphasis on rapid quench detection and minimization of stored magnetic energy.

7. Comparative Assessment and Impact

The integrated SCKansformer model, across all interpretations, prioritizes architectural efficiency, interpretability, and domain-appropriate experimental validation. In image analysis, the KAN-based transformer backbone, redundancy-pruned convolutional modules, and global-local attention mechanisms produce state-of-the-art fine-grained classification results and efficient computation. In tokamak modeling, the exact energy-conserving ODEs and empirical voltage profile closure yield rapidly computable, control-friendly plasma response models. In superconducting transformer contexts, the SCT’s low-resistance, high-coupling architecture enables high-fidelity, safe delivery of extreme currents for magnet technology development.

The term "SCKansformer" thus encompasses a set of rigorously defined, application-matched schemes central to current research in medical image analysis, plasma control engineering, and superconducting instrumentation, each validated through both theoretically grounded construction and empirical protocol (Chen et al., 2024, Romero et al., 2012, Yu et al., 2023).