LUCID-PaTH: Adaptive Ensemble for Spatial Classification

• The paper introduces LUCID-PaTH, an ensemble framework that leverages locally adaptive classifiers and distance-weighted training for classifying complex point sets.
• It utilizes spatial domain adaptation to model variability in cell-type arrangements across tissue regions, providing post-hoc interpretability through discriminative motifs.
• Performance gains over global models are evidenced by improved accuracy and F1-scores on MxIF oncology datasets using multiple ensemble strategies.

LUCID-PaTH (“Locally-Adaptive, Spatially-Lucid Point-set Training and ensembling in non-Euclidean space”) is a spatial-variability–aware, ensemble-based framework for the classification of multi-category point sets in non-Euclidean space. Specifically designed for high-dimensional biomedical imaging tasks—such as classifying multiplexed immunofluorescence (MxIF) maps that encode diverse cell-type locations in tumor tissue—LUCID-PaTH explicitly models both local geometric arrangements of point types and domain shifts across heterogeneous tissue subregions, termed place-types. The framework advances beyond one-size-fits-all deep neural network (DNN) models by introducing locally adaptive classifiers, distance-weighted training, and spatial domain adaptation, all while providing post-hoc spatial lucidity by surfacing the most discriminative cell-type spatial motifs for domain experts (Farhadloo et al., 2024).

1. Problem Motivation and Conceptual Foundations

Modern DNNs for point clouds, such as PointNet and DGCNN, assume invariant spatial structure and are typically trained as global models on all available data. In oncology data derived from MxIF, however, spatial arrangements of cells (e.g., tumor cells, CD8 T cells, macrophages, vasculature) are not globally stationary; they vary dramatically across tissue subregions (“place-types”) such as tumor core, tumor–normal interface, and normal stroma. The same spatial configuration of cell types can have contrasting biological significance depending on the region. Pathologists and immunotherapy researchers require models that not only discern these nuanced spatial relationships but also provide interpretability (“spatial lucidity”) regarding which geometric motifs drive the diagnostic or prognostic classification.

LUCID-PaTH is designed to address two fundamental challenges:

• Spatial variability: Discriminative point-set patterns shift across distinct place-types, invalidating global spatial assumptions.
• Interpretability: Beyond accurate classification, the framework must expose interpretable explanations in terms of underlying k-way local cell-type interactions.

To tackle both, LUCID-PaTH (a) trains separate but connected classifiers for each place-type, (b) leverages a domain expert-defined distance matrix to guide parameter sharing and adaptation, and (c) employs feature permutation-based methods to reveal cell-type spatial motifs after training (Farhadloo et al., 2024).

2. Architecture and Computational Workflow

Input to LUCID-PaTH comprises a collection of $N$ multi-category point sets $X = \{X_i\}$, where each $X_i = \{(x_{ij}, \ell_{ij})\}_j$ encodes cell category $x_{ij}$ and spatial location $\ell_{ij} \in \mathbb{R}^2$. Each point set is annotated with a place-type label $$p_i \in \{\mathrm{PT}_1=\mathrm{normal}, \mathrm{PT}_2=\mathrm{interface}, \mathrm{PT}_3=\mathrm{tumor}\}$$. Domain knowledge of semantic proximity among place-types is encoded in a distance matrix $D = (d_{pq})$, with a threshold $\alpha$ restricting sharing to sufficiently close domains.

The computational pipeline proceeds as follows:

Step	Description	Details
Graph Construction	Build a $k$-NN graph $G_i = (V_i, E_i)$ over each $X_i$ in Euclidean coordinate space	Nodes: one-hot cell type, location; Edges: learned $\alpha_{x_s x_u}$
Base Classifiers	For each place-type $p$, define neural net $h^p$ with $L$ layers of point-wise or edge convolutions	Layer update: $$h_s^{(k+1)}(a,p) = \sigma(W_k^p \sum_{u \in \mathcal N(s)} \alpha_{x_s x_u} h_u^{(k)}(a,p) + B_k^p h_s^{(k)}(a,p))$$
Ensemble Modes	Three training strategies: P1 (separate by place-type), P2 (Weighted-distance Learning Rate, WDLR), P3 (Spatial Domain Adaptation, SDA)	P2: Use $\eta_{p,q} = \eta_0 / d_{pq}$; P3: freeze layers $1\ldots k$, fine-tune $k+1\ldots L$ with MMD loss
Prediction	Place-type–specific or weighted ensemble soft-labeling	Aggregate predictions with weights $\propto 1/d_{pq}$

This approach allows LUCID-PaTH to model both within-region and across-region spatial discriminative features.

3. Mathematical Formulation of Training and Adaptation

3.1 Weighted-Distance Learning Rate (WDLR)

Given a training example $(X_j, y_j, p_j)$ and a classifier $h_{p_i}$ for place-type $p_i$, the framework assigns a distance-weighted sample weight

$w_{ij} = f(d_{p_i p_j}) = 1 / d_{p_i p_j}$

The per-sample learning rate is

$\eta_{ij} = \eta_0 \cdot w_{ij}$

The weighted cross-entropy objective becomes:

$$L_{WDLR}(\theta; p_i) = \sum_{j: d_{p_i p_j} \leq \alpha} w_{ij} \cdot \ell(h_{p_i}(X_j; \theta), y_j)$$

3.2 Spatial Domain Adaptation (SDA)

Parameter vector $\theta$ is split as $[\theta_k \| \theta_{k+1:\,L}]$. Fixing the first $k$ layers on all source place-types $S$, for target $T=\{p\}$, SDA minimizes: $$\min_{\theta_{k+1:L}'} \sum_{(x_t, y_t) \in D_T} \ell(h(x_t; \theta_k, \theta_{k+1:L}'), y_t) + \lambda D_{\mathrm{MMD}}(\{h(x_s; \theta_k)\}_{s \in D_S}, \{h(x_t; \theta_k)\}_{t \in D_T})$$ Here, $D_{\mathrm{MMD}}$ denotes the Maximum Mean Discrepancy between the feature distributions at the frozen layer (Farhadloo et al., 2024).

3.3 Joint Optimization

Combining WDLR and SDA in training yields: $$\min_{\theta} \sum_i L_{WDLR}(\theta;p_i) + \lambda \mathcal{L}_{adapt}(\theta)$$

4. Place-types, Handling Spatial Variability, and Data Protocol

Place-types are formalized as spatial domains $\mathscr X_p$ with their own point-set distributions $P_p(X)$. Assignment of tissue regions to place-types (normal, interface, tumor) is determined by pathology rules, and the domain expert–driven distance matrix $d_{pq}$ encodes contextual similarity among regions.

Training protocols employ an 80%/20% split (with 25% of the training set as validation), horizontal MBR partitions, multiple rotations for augmentation, and uniform downsampling to 1,024 points per view. Supported architectures include PointNet, DGCNN, Point Transformer, and SAMCNet. Hyperparameter settings recommended for LUCID-PaTH include a base learning rate $\eta_0 = 1 \mathrm{e}{-3}$, $\alpha = 3$ (so WDLR rates of $1\mathrm{e}{-3}$, $5\mathrm{e}{-4}$, $3.3\mathrm{e}{-4}$ for distances 1 to 3), batch size 32, Adam optimizer ($\beta_1=0.9$, $\beta_2=0.999$), and freezing the first $k=2$ layers in SDA (Farhadloo et al., 2024).

5. Quantitative Performance and Empirical Insights

Experiments conducted on MxIF oncology datasets stratified into three place-types (PT$_1$: normal, PT$_2$: interface, PT$_3$: tumor; sample counts: 81, 145, 103 fields of view) compared LUCID-PaTH ensemble strategies against a one-size-fits-all (OSFA) baseline. Summary metrics (weighted average Accuracy, F1-score) for the SAMCNet backbone are as follows:

Method	Accuracy	F1-score
OSFA	0.714	0.714
P1 (sep.)	0.857	0.857
P2 (WDLR)	0.806	0.806
P3 (SDA)	0.824	0.856

Across all architectures tested, P1 and P3 consistently outperform the global baseline by 7–14% accuracy. Notably, PointNet+WDLR and Point Transformer+SDA each resulted in +7% accuracy over OSFA, and LUCID-PaTH with SAMCNet produces an overall +33.6% accuracy increase when ensemble and domain adaptation are combined (Farhadloo et al., 2024).

6. Model Interpretability and Domain Insights

Post-training, LUCID-PaTH applies permutation-based feature importance analysis to learned SAMCNet embeddings, surfacing the top-ranked spatial motifs responsible for predictions within each place-type. Illustrative results include:

• OSFA (entire tissue): (Vasculature–Helper T–Macrophage), (Helper T–Macrophage–Tumor–Vasculature), etc.
• Tumor (PT$_3$): (Tumor–Tumor–Vasculature), (Macrophage–Tumor), (Tumor–Macrophage–Vasculature)
• Interface (PT$_2$): (B cell–Helper T–Vasculature), (B cell–Helper T–Tumor), (B cell–Macrophage–Reg T)

These discriminative motifs align with known biological mechanisms: for instance, angiogenesis and tumor–macrophage interaction dominate tumor regions, while B cell–T cell–vasculature co-location characterizes interface zones where lymphoid aggregation and antibody-mediated responses emerge. The spatial lucidity of LUCID-PaTH enables pathologists to directly inspect which k-way cell co-locations underlie the predicted immunotherapy response (Farhadloo et al., 2024).

7. Current Limitations and Prospective Extensions

LUCID-PaTH, while effective for inter-place-type variability, does not yet directly address finer granularities such as necrotic core vs. hypoxic rim, nor does it include temporal (spatio-temporal) modeling. Anticipated extensions include: (a) dimension-adaptive graphs or attention mechanisms for more complex MxIF panels, (b) generative models such as GANs to synthesize rare spatial phenotypes and augment low-frequency classes, and (c) explicit modeling of temporal dynamics in point-set sequences. This suggests further generalizations of LUCID-PaTH could address these facets and broaden its applicability across biomedical and spatial analytics domains (Farhadloo et al., 2024).