arXiv subjects

Edward Castillo

Publications and source records attributed to Edward Castillo.

5 recordsLinked to original sources

arXiv2026

ConceptM$^3$oE: Concept-Guided Multimodal Mixture of Experts for Interpretable Computational Pathology

Healthcare models are transitioning from unimodal prediction toward multimodal reasoning over heterogeneous diagnostic inputs. In computational pathology, for complex tumor subtypes where morphology alone can be challenging to distinguish, pathology reports and molecular measurements may provide additional diagnostic evidence alongside whole-slide images, yet existing models often fail to clarify how diverse signals assemble into recognizable diagnostic concepts. We propose ConceptM$^3$oE (Concept Multimodal MoE), which embeds concept formation directly within interaction-aware mixture-of-experts (MoE) pathways. The architecture decomposes evidence into modality-specific, redundant, and synergistic experts, which are then projected into structured concept bottlenecks mapping latent features to a hierarchy of morphology and biomarker concepts. To prevent the information loss typical of interpretable bottlenecks, we utilize residual pathways within each expert to allow task-relevant signals to flow both through the concepts and directly to the final task prediction, so that high performance is maintained alongside interpretability. Across an institutional pediatric brain tumor cohort and a public glioma cohort, the framework delivers competitive performance to unconstrained models while producing reasoning traces validated by an independent neuropathologist. In data-limited regimes, ConceptM$^3$oE improves limited-data performance, increasing macro-F1 from 56.41% to 66.70% at small training sizes compared to non-concept-informed baselines, while also showing faster training convergence consistent with the regularizing effect of concept learning. This work offers a scalable path toward high-performance medical AI that is inherently verifiable and better aligned with the complex decision-making of clinical practice.

cs.AI↗

arXiv2026

PathMoE: Interpretable Multimodal Interaction Experts for Pediatric Brain Tumor Classification

Accurate classification of pediatric central nervous system tumors remains challenging due to histological complexity and limited training data. While pathology foundation models have advanced whole-slide image (WSI) analysis, they often fail to leverage the rich, complementary information found in clinical text and tissue microarchitecture. To this end, we propose PathMoE, an interpretable multimodal framework that integrates H\&E slides, pathology reports, and nuclei-level cell graphs via an interaction-aware mixture-of-experts architecture built on state-of-the-art foundation models for each modality. By training specialized experts to capture modality uniqueness, redundancy, and synergy, PathMoE employs an input-dependent gating mechanism that dynamically weights these interactions, providing sample-level interpretability. We evaluate our framework on two dataset-specific classification tasks on an internal pediatric brain tumor dataset (PBT) and external TCGA datasets. PathMoE improves macro-F1 from 0.762 to 0.799 (+0.037) on PBT when integrating WSI, text, and graph modalities; on TCGA, augmenting WSI with graph knowledge improves macro-F1 from 0.668 to 0.709 (+0.041). These results demonstrate significant performance gains over state-of-the-art image-only baselines while revealing the specific modality interactions driving individual predictions. This interpretability is particularly critical for rare tumor subtypes, where transparent model reasoning is essential for clinical trust and diagnostic validation.

cs.CV↗

arXiv2025

EVO-LRP: Evolutionary Optimization of LRP for Interpretable Model Explanations

Explainable AI (XAI) methods help identify which image regions influence a model's prediction, but often face a trade-off between detail and interpretability. Layer-wise Relevance Propagation (LRP) offers a model-aware alternative. However, LRP implementations commonly rely on heuristic rule sets that are not optimized for clarity or alignment with model behavior. We introduce EVO-LRP, a method that applies Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to tune LRP hyperparameters based on quantitative interpretability metrics, such as faithfulness or sparseness. EVO-LRP outperforms traditional XAI approaches in both interpretability metric performance and visual coherence, with strong sensitivity to class-specific features. These findings demonstrate that attribution quality can be systematically improved through principled, task-specific optimization.

cs.LG↗

arXiv2018

Deriving ventilation imaging from 4DCT by deep convolutional neural network

Purpose: Functional imaging is emerging as an important tool for lung cancer treatment planning and evaluation. Compared with traditional methods such as nuclear medicine ventilation-perfusion (VQ), positron emission tomography (PET), single photon emission computer tomography (SPECT), or magnetic resonance imaging (MRI), which use contrast agents to form 2D or 3D functional images, ventilation imaging obtained from 4DCT lung images is convenient and cost-effective because of its availability during radiation treatment planning. Current methods of obtaining ventilation images from 4DCT lung images involve deformable image registration (DIR) and a density (HU) change-based algorithm (DIR/HU); therefore the resulting ventilation images are sensitive to the selection of DIR algorithms. Methods: We propose a deep convolutional neural network (CNN)-based method to derive the ventilation images from 4DCT directly without explicit DIR, thereby improving consistency and accuracy of ventilation images. A total of 82 sets of 4DCT and ventilation images from patients with lung cancer were studied using this method. Results: The predicted images were comparable to the label images of the test data. The similarity index and correlation coefficient averaged over the ten-fold cross validation were 0.883+-0.034 and 0.878+-0.028, respectively. Conclusions: The results demonstrate that deep CNN can generate ventilation imaging from 4DCT without explicit deformable image registration, reducing the associated uncertainty.

physics.med-ph↗

arXiv2009

Implementation and evaluation of various demons deformable image registration algorithms on GPU

Online adaptive radiation therapy (ART) promises the ability to deliver an optimal treatment in response to daily patient anatomic variation. A major technical barrier for the clinical implementation of online ART is the requirement of rapid image segmentation. Deformable image registration (DIR) has been used as an automated segmentation method to transfer tumor/organ contours from the planning image to daily images. However, the current computational time of DIR is insufficient for online ART. In this work, this issue is addressed by using computer graphics processing units (GPUs). A grey-scale based DIR algorithm called demons and five of its variants were implemented on GPUs using the Compute Unified Device Architecture (CUDA) programming environment. The spatial accuracy of these algorithms was evaluated over five sets of pulmonary 4DCT images with an average size of 256x256x100 and more than 1,100 expert-determined landmark point pairs each. For all the testing scenarios presented in this paper, the GPU-based DIR computation required around 7 to 11 seconds to yield an average 3D error ranging from 1.5 to 1.8 mm. It is interesting to find out that the original passive force demons algorithms outperform subsequently proposed variants based on the combination of accuracy, efficiency, and ease of implementation.

physics.med-ph↗