XMorph: Medical Imaging & Robotics Frameworks

• XMorph is a dual framework system where one branch focuses on explainable brain tumor classification using hybrid feature fusion and the other on translating human motion to robot behavior.
• The medical imaging framework integrates nonlinear boundary analysis, fractal and chaotic descriptors, and dual-channel explainability to achieve ~96% classification accuracy with modest computation.
• The robotics framework employs cross-morphology retargeting, physics-aware correction, and privileged reinforcement learning to convert human motion into deployable policies for diverse non-humanoid robots.

XMorph denotes two distinct 2026 research frameworks that share closely related names but address different technical problems. In medical imaging, "XMorph: Explainable Brain Tumor Analysis Via LLM-Assisted Hybrid Deep Intelligence" defines an explainable and computationally efficient framework for fine-grained classification of glioma, meningioma, and pituitary tumors from MRI, centered on Information-Weighted Boundary Normalization (IWBN), hybrid feature fusion, and dual-channel explainability (Ghahfarokhi et al., 24 Feb 2026). In robotics, "X-Morph: Human Motion Priors for Scalable Robot Learning Across Morphologies" defines a human-motion-to-robot-behavior pipeline that converts human motion into deployable locomotion and loco-manipulation policies for diverse non-humanoid legged morphologies through cross-morphology retargeting, privileged reinforcement learning, and causal student distillation (Sharma et al., 29 Jun 2026).

1. Nomenclature and domain separation

A recurrent source of confusion is purely terminological: XMorph and X-Morph are not the same system. They are separate frameworks introduced in different application domains and with different technical objectives (Ghahfarokhi et al., 24 Feb 2026, Sharma et al., 29 Jun 2026).

Name	Domain	Core objective
XMorph	Brain tumor MRI analysis	Fine-grained classification of glioma, meningioma, and pituitary tumors with explainability and modest computational demands
X-Morph	Robot learning across morphologies	Convert human motion into deployable locomotion and loco-manipulation policies for non-humanoid legged robots

The medical-imaging XMorph is organized around segmentation, tumor-specific morphology descriptors, deep feature extraction, XGBoost classification, and a dual-channel explainable AI module. The robotics X-Morph is organized around retargeting, physics-aware correction, privileged RL tracking, student-policy distillation, and real-time causal retargeting. This separation is essential because the shared name does not imply shared methods, datasets, or evaluation criteria.

2. XMorph in brain tumor analysis: motivation and problem formulation

The medical XMorph is motivated by both clinical and technical constraints. Early and accurate classification of glioma, meningioma, and pituitary tumors via MRI is critical for treatment planning and prognosis, while clinicians require transparent reasoning—such as why a lesion is classified as glioma—to trust AI-based decisions in high-stakes settings (Ghahfarokhi et al., 24 Feb 2026). At the technical level, deep CNNs are described as “black boxes,” and large, over-parameterized models such as MGMT-net incur seconds of per-slice inference, which is unsuitable for real-time workflows or resource-limited devices. The framework is also motivated by the observation that tumor boundaries exhibit nonlinear, fractal-like irregularities associated with infiltrative growth and that these are not fully captured by standard CNN features.

XMorph therefore defines three explicit goals. First, it integrates nonlinear boundary metrics, including chaotic and fractal descriptors, together with quantitative clinical biomarkers into a lightweight and interpretable pipeline. Second, it provides dual-channel explainability, combining a visual channel and a textual channel that link model outputs to clinically meaningful features. Third, it aims for competitive accuracy of approximately 96% while maintaining modest computational demands.

The problem setting is fine-grained multiclass classification over three prominent brain tumor types. The framework is not posed as end-to-end raw-image classification alone; instead, it is a hybrid system in which segmentation-derived morphology, handcrafted nonlinear descriptors, clinical biomarkers, and deep visual embeddings are fused before classification. A plausible implication is that the design treats morphological irregularity as diagnostically central rather than merely auxiliary.

3. XMorph architecture for MRI-based classification

XMorph processes an input MRI through six stages (Ghahfarokhi et al., 24 Feb 2026).

Stage 1: automated tumor segmentation. Tumor segmentation is performed with DeepLabV3 using a ResNet-50 backbone and a combined cross-entropy plus Dice loss,

$$L_{\mathrm{combined}} = L_{\mathrm{CE}} + L_{\mathrm{Dice}}.$$

Training uses data augmentations consisting of $\pm 15^\circ$ rotations, flips, and intensity scaling. On the reported segmentation task, the model achieved overall Dice $\approx 0.932$.

Stage 2: tumor-specific feature extraction. This stage begins with boundary-to-signal conversion. A 2D tumor contour is extracted and resampled to $N=256$ points $(x_i,y_i)$. With centroid $(c_x,c_y)$, radial distances are defined as

$$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$

and converted into the scale-invariant normalized signal

$$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$

From this representation, geometric indices are computed:

• Irregularity Index $= \sigma(S_{\mathrm{std}})$
• Roughness Index $$= \sum_i |S_{\mathrm{std},i+1} - S_{\mathrm{std},i}|$$

The central morphological mechanism is Information-Weighted Boundary Normalization (IWBN). At each boundary point, XMorph computes local entropy $\pm 15^\circ$0, for example sample entropy over a curvature window, then normalizes entropy as

$\pm 15^\circ$1

Information weights are defined by

$\pm 15^\circ$2

and the IWBN signal is

$\pm 15^\circ$3

Derived indices are the Mean Local Entropy

$\pm 15^\circ$4

the Weight Range

$\pm 15^\circ$5

and the Enhancement Factor

$\pm 15^\circ$6

XMorph further augments this representation with nonlinear and chaotic features, including fractal dimension by box-counting,

$\pm 15^\circ$7

Approximate Entropy, Sample Entropy, Permutation Entropy, and the Largest Lyapunov Exponent,

$\pm 15^\circ$8

It also includes quantitative clinical biomarkers derived from ROI and intensity masks:

• Ring Enhancement Index:

$\pm 15^\circ$9

• Skull-to-Tumor Distance:

$\approx 0.932$0

• Midline Shift:

$\approx 0.932$1

Stage 3: deep feature extraction. The full slice is resized to $\approx 0.932$2, passed through a pre-trained ResNet-50, and reduced from a 2048-dimensional GAP output using PCA retaining 95% variance.

Stage 4: hybrid feature fusion. The fused representation is

$\approx 0.932$3

where $\approx 0.932$4 concatenates geometric, IWBN, chaotic, and clinical tumor-specific features.

Stage 5: tumor classification. Classification uses XGBoost with 300 trees, $\approx 0.932$5, and $\approx 0.932$6, trained with multiclass logistic loss and tree-complexity regularization,

$\approx 0.932$7

under 5-fold stratified cross-validation.

Stage 6: dual-channel explainable AI. The visual channel uses GradCAM++ on the last CNN convolutional layer to generate a class-specific heatmap overlaid on MRI. The textual channel uses SHAP decomposition of the XGBoost prediction,

$\approx 0.932$8

selects top-$\approx 0.932$9 features by $N=256$0, forms a structured prompt $N=256$1, sends it to the GPT-5 API, and receives a clinical-style rationale $N=256$2 describing key features and uncertainty. The resulting interface displays the GradCAM++ map and LLM-generated report side by side.

4. Performance, interpretability, and limitations of the medical XMorph

On segmentation, XMorph reports that DeepLabV3 outperformed U-Net, with overall Dice of $N=256$3 versus $N=256$4, IoU of $N=256$5 versus $N=256$6, Precision of $N=256$7 versus $N=256$8, and Recall of $N=256$9 versus $(x_i,y_i)$0 (Ghahfarokhi et al., 24 Feb 2026). On classification under 5-fold cross-validation, Tumor-Specific features only achieved accuracy $(x_i,y_i)$1, sensitivity $(x_i,y_i)$2, and specificity $(x_i,y_i)$3; Deep-features only achieved accuracy $(x_i,y_i)$4, sensitivity $(x_i,y_i)$5, and specificity $(x_i,y_i)$6; and the fused XMorph representation achieved accuracy $(x_i,y_i)$7, sensitivity $(x_i,y_i)$8, and specificity $(x_i,y_i)$9. The macro-average ROC AUC values were 0.92 for Tumor-Specific, 0.95 for Deep, and 0.98 for Fused. The reported ablation result states that nonlinear and clinical descriptors boost deep models by +3% accuracy.

The computational profile is explicitly part of the framework definition. DeepLabV3 segmentation is reported at approximately $(c_x,c_y)$0 s per slice on an NVIDIA GPU; ResNet-50 plus PCA at approximately $(c_x,c_y)$1 s per slice; IWBN, chaotic, and clinical features at approximately $(c_x,c_y)$2 s per slice on CPU; and XGBoost prediction plus XAI at approximately $(c_x,c_y)$3 s. Total inference is approximately $(c_x,c_y)$4 s per slice, described as an order of magnitude faster than heavy 3+ s architectures.

The explainability mechanism combines spatial attribution and semantic rationalization. The visual channel highlights tumor margins, ring-enhancing regions, and midline displacement zones. The textual channel is exemplified by the following rationale: “This lesion is classified as glioma (98% confidence) because it exhibits high local entropy (0.76), an elevated fractal dimension (D=1.42), significant ring enhancement index (0.47), and a midline shift of 13.8%, consistent with infiltrative, high-grade pathology.” In the radiologist use-case, the overlay map guides ROI inspection and the narrative identifies which boundary irregularities and biomarkers drove the decision. This suggests that XMorph is intended not only to improve classification metrics but also to externalize intermediate evidence in clinically legible form.

The reported limitations are equally explicit. The current model is trained on single-center T1-weighted data, and generalization to multi-center, multi-parametric MRI such as T2 and FLAIR is pending. The method depends on segmentation quality, so ROI delineation errors propagate to boundary features. LLM rationales may still hallucinate if SHAP inputs are noisy; safety constraints mitigate but do not eliminate this risk. Proposed future work includes extension to Vision–LLMs for integrated multimodal reasoning, validation on prospective clinical trials, and incorporation of advanced uncertainty quantification.

5. X-Morph in robotics: retargeting, correction, and policy learning

The robotics X-Morph addresses a different problem: abundant human motion data exists, but comparable motion data is scarce for non-humanoid legged robots such as quadrupeds, hexapods, and quadruped manipulators. Direct retargeting is characterized as often producing motions that are visually plausible yet physically inconsistent or difficult to track under robot dynamics. X-Morph therefore defines a pipeline that converts human motion into deployable robot behaviors while preserving intent and respecting morphology-specific constraints (Sharma et al., 29 Jun 2026).

The first stage is cross-morphology retargeting. Given a human motion sequence

$(c_x,c_y)$5

the goal is to produce robot joint-space references

$(c_x,c_y)$6

for each robot morphology $(c_x,c_y)$7, preserving root velocity, end-effector trajectories, and contact timing. The mapping

$(c_x,c_y)$8

is implemented as a neural network $(c_x,c_y)$9. At frame $$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$0, the network predicts

$$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$1

where $$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$2 are target joint angles, $$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$3 is the local-heading linear velocity, and $$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$4 is the yaw rate. Root pose is reconstructed by integrating $$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$5 and $$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$6 over time.

Retargeter training uses a composite objective,

$$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$7

containing reconstruction, adversarial, latent cycle-consistency, motion cycle-consistency, forward-kinematics cycle, velocity, yaw, joint-limit, foot-skate, grounding, and end-effector correspondence terms. End-effector loss is defined through morphology-aware forward kinematics:

$$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$8

Joint limits are softly enforced by

$$r_i = \sqrt{(x_i-c_x)^2 + (y_i-c_y)^2}, \quad i=1,\dots,N,$$9

and foot skating and grounding are defined with contact bodies and a nominal ground plane at $$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$0:

$$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$1

$$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$2

The framework states that no explicit iterative inverse-kinematics solver is run at test time; instead, the learned retargeter amortizes this optimization.

Because the retargeter may still produce contact artifacts, X-Morph introduces a physics-aware offline correction network $$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$3 operating on full clips. It predicts residual corrections $$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$4, $$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$5, and $$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$6 so that

$$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$7

This corrector is trained with additional losses on foot penetration, skating, grounding, joint limits, and temporal smoothness.

Tracking these references is posed as a privileged reinforcement-learning problem. For each morphology $$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$8, the MDP uses privileged state information containing proprioceptive state, reference context, and privileged signals such as base-height error, joint torques, contact flags, and motion identity. Actions are desired joint positions for a PD controller. The reward is a weighted sum of imitation and physical-penalty terms:

$$S_{\mathrm{std},i} = \frac{r_i}{\left(\sum_i r_i / N\right)}.$$9

To accelerate learning, X-Morph adds the APEX action prior:

$= \sigma(S_{\mathrm{std}})$0

with $= \sigma(S_{\mathrm{std}})$1 over training. After the teacher converges, a causal student policy is distilled from teacher actions using

$= \sigma(S_{\mathrm{std}})$2

or alternatively a KL term if the teacher is stochastic. For deployment, X-Morph also trains a causal retargeter $= \sigma(S_{\mathrm{std}})$3 for real-time operation from short causal history.

6. Robotics evaluation, throughput, and downstream uses

X-Morph is evaluated on three platforms: the Go2 quadruped, the Yuna hexapod, and the B2-Z1 quadruped equipped with a 2-DoF manipulator (Sharma et al., 29 Jun 2026). The reported reference-quality ablation on Go2 locomotion over 33 clips shows that adding the offline corrector $= \sigma(S_{\mathrm{std}})$4 reduces foot slip by 27.2% from 58.8 to 42.8 cm/s, foot penetration p95 by 46.9% from 11.34 to 6.02 cm, contact-height error by 44.2% from 6.45 to 3.60, floating error by 39.3% from 3.09 to 1.88, and joint-acceleration p95 by 13.9% from 32.61 to 28.08 rad/s$= \sigma(S_{\mathrm{std}})$5.

Under live video references on Yuna, training on corrected references improves joint-MAE from 6.57 to 5.45 degrees, root-velocity RMSE from 0.479 to 0.413 m/s, yaw-rate RMSE from 0.896 to 0.651 rad/s, base-height standard deviation from 2.12 to 1.72 cm, and foot slip from 29.29 to 24.30 cm/s. For video teleoperation throughput, the live pipeline publishes references on a 30 Hz camera stream at up to 24.3 Hz without visualization and 20.0 Hz with visualization in sim2sim tests, and 28.9 Hz peak on real hardware.

Qualitative behavior coverage includes forward and backward walking, turning, squatting, and expressive arm gestures on Go2 and Yuna; loco-manipulation such as box picking from human arm motions on Yuna; and front-arm reaching for a door handle on B2-Z1. For downstream task initialization, retargeted human door-opening motions on Yuna are used as a structured prior, after which a subsequent RL policy with no further human data learns to open a door from this initialization far more stably than from scratch, reported qualitatively.

The downstream applications are organized around three interfaces. In video-based teleoperation, the pipeline is monocular RGB to FastSAM3D Body to GMR to G1 motion to causal retargeter to student policy, achieving approximately 20 to 29 Hz reference updates. In text-conditioned motion control, prompts such as “Pick up box” are converted by Kimodo into G1 motion, then passed through retargeting and policy execution; the same human-motion library is reused without task-specific training for each morphology. In behavior priors for RL, retargeted human motions provide structured exploration behaviors such as walking and reaching, demonstrated on hexapod door opening and stated to generalize to other loco-manipulation domains. Additional conditioning can be applied to both offline and causal retargeters via concatenated embeddings in the body-part latent space.

Taken together, the two frameworks named XMorph/X-Morph instantiate markedly different research programs. The medical XMorph uses morphology-sensitive boundary analysis, deep visual features, and LLM-assisted explanation to support brain tumor classification, whereas the robotics X-Morph uses human motion priors, morphology-aware retargeting, and RL distillation to scale behavior learning across robot embodiments. The shared naming convention obscures this distinction; technically, they are independent systems with different data regimes, optimization objectives, and deployment contexts (Ghahfarokhi et al., 24 Feb 2026, Sharma et al., 29 Jun 2026).