A little about me
I’m a graduate student in the Computer Science department at Marquette University , working at the intersection of machine learning, computational pathology, and model interpretability.
My interests include foundation models, unsupervised pre-training, generative AI, large language models, and adversarial attacks. I’m drawn to the question of what models learn—and how we can make sense of it.
Previously, I was a lecturer in the Electrical, Electronic, and Communication Engineering department at MIST . Teaching continues to shape how I approach research: work through the details, connect the ideas, and share what I learn.
Current chapter
Graduate studies Marquette University
Research focus
Vision & interpretable AI Foundation models · Computational pathology
Follow my experience
I study multimodal and agentic AI in computational pathology. My research explores how we can make these systems interpretable, analyze uncertainty in their predictions, understand their vulnerabilities, and learn from complex data.
Computational pathology Explainable AI Graph learning
Dibaloke Chanda , Nasim Yahya Soltani, Masoud Ganji
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)Under review
CThoughtPathNet: Expert-Guided Vision–Language Reasoning for Few-Shot Cancer Subtyping in Whole-Slide Pathology
Journal · 2027Under review
Expert-guided vision–language reasoning for few-shot cancer subtyping in whole-slide pathology.
Big idea: Use domain-specific expert roles and model-generated reasoning to guide few-shot cancer subtyping from whole-slide images.Method: Combine four expert-role prompts and reasoning-informed text embeddings with 10×/20× visual features using learned patch prototypes and cross-attention.Finding: Under the 16-shot, five-fold protocol, the best reported configurations achieve AUC 0.9604 ± 0.0088 on TCGA-RCC and 0.8770 ± 0.0195 on TCGA-NSCLC (mean ± standard deviation), outperforming the 20 evaluated MIL baselines.Impact: Shows how expert-derived language and multi-scale visual evidence can support cancer subtyping with limited labeled slides; benefits depend on the chosen encoder and prompting configuration.
Authors Dibaloke Chanda , Nasim Yahya Soltani, Masoud Ganji
Publication IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)
Research themes Computational pathology Vision–language reasoning Few-shot learning Cancer subtyping
Dibaloke Chanda , Praveen Madiraju
IEEE International Conference on Big Data (IEEE BigData)Accepted
INVEX: Label-Free Pathology Foundation Model Ranking via Invariance and Expressiveness
Conference · 2026Accepted
Ranking pathology foundation models without labels using invariance and expressiveness.
Big idea: Rank pathology foundation models without labels, fine-tuning, or training a downstream predictor.Method: Combine rotational invariance with representation expressiveness, using a standardized score based on neighborhood retention and effective rank.Finding: Across 22 encoders and 16 datasets containing over 1.2 million patches, ranking agreement reaches Spearman correlations of 0.749 with a linear oracle and 0.683 with a kNN oracle.Impact: Provides an annotation-free way to shortlist encoders before investing in task-specific evaluation.
Authors Dibaloke Chanda , Praveen Madiraju
Publication IEEE International Conference on Big Data (IEEE BigData)
Research themes Foundation models Model ranking Label-free evaluation Computational pathology
Dibaloke Chanda , Praveen Madiraju
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)Accepted
MOSAIC: Towards a Universal Latent Space for Computational Pathology Foundation Models
Conference · 2027Accepted
Exploring a universal latent space for computational pathology foundation models.
Big idea: Test whether independently trained pathology models can share a useful latent representation despite incompatible native embedding spaces.Method: Compare 18 patch- and slide-level encoders across 4,331 TCGA/CPTAC slides, using similarity analysis, multiview alignment, retrieval, and 13 downstream tasks.Finding: Shared-space features improve selected molecular tasks, including BRCA GATA3 AUC from approximately 0.55 to 0.63, while underperforming on some other targets.Impact: Supports cross-model interoperability while showing that geometric alignment does not guarantee better performance on every pathology task.
Authors Dibaloke Chanda , Praveen Madiraju
Publication IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)
Research themes Computational pathology Foundation models Latent representations
Dibaloke Chanda , Nasim Yahya Soltani, Masoud Ganji
Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
MERI-VL-MIL: Multi-Expert Reasoning-Informed Vision-Language Multi-Instance Learning for Few-Shot Cancer Subtyping
Conference · 2026
Multi-expert reasoning-informed vision–language multi-instance learning for few-shot cancer subtyping.
Big idea: Use multiple expert perspectives and model-generated reasoning to enrich language guidance for few-shot cancer subtyping.Method: Prompt three multimodal language models as four diagnostic experts, encode their responses and reasoning, and fuse the pooled embeddings with 10×/20× whole-slide features.Finding: On TCGA lung subtyping, GPT-5 with mean pooling reports 71.6% mean accuracy (standard deviation 23.1 percentage points) across five runs; pooled experts outperform individual experts.Impact: Shows how complementary expert-derived language can guide visual multiple-instance learning when labeled slides are limited.
Authors Dibaloke Chanda , Nasim Yahya Soltani, Masoud Ganji
Publication Proceedings of the IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)
Research themes Multi-instance learning Vision–language models Few-shot learning Cancer subtyping
Dibaloke Chanda , Milan Aryal, Masoud Ganji, Anil Parwani, Nasim Yahya Soltani
Journal of Pathology Informatics
A New Era in Computational Pathology: A Survey on Foundation and Vision-Language Models
Journal · 2026
A survey of foundation models and vision-language models in computational pathology.
Big idea: Map the shift from task-specific pathology models toward reusable foundation and vision-language models.Method: Organize the literature into vision foundation models, vision-language models, multimodal datasets, and benchmarking and adaptation approaches.Synthesis: Connects model development with supporting datasets, evaluation frameworks, prompting, and vision-instruction tuning in a unified research timeline.Impact: Provides a guide for navigating pathology models and datasets and understanding how the research directions fit together.
Authors Dibaloke Chanda , Milan Aryal, Masoud Ganji, Anil Parwani, Nasim Yahya Soltani
Publication Journal of Pathology Informatics
Research themes Computational pathology Foundation models Vision-language models
Dibaloke Chanda , Saba Heidari Gheshlaghi, Nasim Yahya Soltani
Knowledge-Based Systems
Explainability-Based Adversarial Attack on Graphs Through Edge Perturbation
Journal · 2025
An explainability-based approach to adversarial attacks on graphs through edge perturbation.
Big idea: Explanations of a graph neural network's predictions can reveal which parts of the graph an adversary should target.Method: Use GNNExplainer and PGExplainer to identify important subgraphs, then perturb graph edges to construct adversarial test samples.Finding: Explanation-guided attacks can avoid complex attack optimization and training-time model access by exploiting information from a trained model.Impact: Reveals a security tradeoff in model transparency and motivates evaluating explainability alongside adversarial robustness.
Authors Dibaloke Chanda , Saba Heidari Gheshlaghi, Nasim Yahya Soltani
Publication Knowledge-Based Systems
Research themes Explainability Adversarial attacks Graph learning Edge perturbation
Dibaloke Chanda , Nasim Yahya Soltani
IEEE Transactions on Power Systems
A Heterogeneous Graph-Based Multi-Task Learning for Fault Event Diagnosis in Smart Grid
Journal · 2025
Heterogeneous graph-based multi-task learning for diagnosing fault events in smart grids.
Big idea: Diagnose multiple aspects of a smart-grid fault together using the network's graph structure.Method: A shared heterogeneous graph encoder feeds task-specific classification and regression heads, trained with a combined multi-task objective.Outputs: Jointly addresses fault detection, fault-type classification, and localization alongside fault-resistance and current estimation.Impact: Connects whether, what, and where a fault occurs with estimates of its electrical characteristics for a more complete diagnosis.
Authors Dibaloke Chanda , Nasim Yahya Soltani
Publication IEEE Transactions on Power Systems
Research themes Heterogeneous graphs Multi-task learning Fault diagnosis Smart grids
Retrieval-augmented generation
An interactive map of vanilla RAG, GraphRAG variants, and agentic retrieval, with stage-by-stage explanations, worked mathematics, and code examples.
RAG GraphRAG Agents
Explore project
Language model internals
A visual guide to the design choices inside language models. Explore the forward pass, compare architecture families, and trace real model configurations through the design space.
LLMs Transformers Architecture
Explore project
Reinforcement learning
An interactive explorer of reinforcement learning concepts and algorithm families, bringing together a navigable taxonomy, equations, notation, and visual comparisons.
Reinforcement learning Algorithms Equations
Explore project
Computational pathology
An interactive catalog of pathology foundation models, organized by model family and application, with summaries of pretraining data, key ideas, and links to papers, code, and model releases.
Foundation models Vision-language Model taxonomy
Explore project
Agents & computational pathology
An interactive catalog of pathology agents and copilots, covering diagnostic reasoning, slide navigation, multi-agent systems, tool use, and evaluation, with links to the underlying research.
Agentic AI Computational pathology Research catalog
Explore project
Whole-slide pathology benchmarking
200 TB+ of WSI data processed, with inventories spanning 39 WSI datasets and 20 patch-level datasets across diverse organs and pathology tasks.
200 TB+ processed 39 WSI datasets 20 patch datasets
Explore project
Learning in public
Ideas worthsharing. I write about deep learning, mathematics, and the things I learn along the way. My notes connect the intuition with the details.
Explore my writing Deep learning Linear algebra Machine learning
