Ashim Dhor

news

• October 2026 Paper "Is the Circuit in the Model or in the Method? Spectral Identifiability for Mechanistic Evaluation" accepted at three NeurIPS 2026 workshops: Can We Trust AI Evaluation?, XAI4Science: Knowledge Discovery and Trust through Interpretable Foundation Models and NewInML.
• July 2026 Abstract accepted for Oral Presentation at the Asian Society of Digital Pathology (ASDP) 2026, Singapore, and received a grant to present my work at the Congress.
• July 2026 Selected for the NeurIPS 2026 Affinity Support Pilot Program.
• May 2026 Selected as a Reviewer for Trustworthy AI for Good, ICML 2026 Workshop .
• March 2026 Paper MOSAIC accepted at IEEE ICHI 2026 Young Scholar Symposium. Selected for the Rising Star Award.
• February 2026 Our paper TUNE++ on topology-guided uncertainty for 3D medical segmentation is accepted at MIDL 2026.
• January 2026 Paper on HISTO-UNet for histopathology segmentation accepted at IEEE ISBI 2026.
• 2026 Awarded the MICCAI Society Membership Grant 2026.
• November 2025 Received Student Scholarship Travel Grant of USD 1100 for presenting my work at AAAI 2026.
• October 2025 AAAI 2026 Undergraduate Consortium Mentor.
• October 2025 Topo-GraT accepted at AAAI 2026 (Student Abstract Track).
• 2025 Selected as a Reviewer for CHIL 2025, ML4H 2025, and AAAI 2026.
• 2025 Started as Bioinformatics Scientist at Elucidata.
• 2025 Selected for NBEC 2025 second round for HistoAI, an AI-powered cancer diagnostic tool concept.
• 2024 Paper on Financial Fraud Detection using Quantum GNNs published in Quantum Machine Intelligence.

publications

2026

TUNE++: Topology-Guided Uncertainty Estimation for Reliable 3D Medical Image Segmentation

Ashim Dhor, Abhirup Banerjee, Tanmay Basu

Medical Imaging with Deep Learning (MIDL) 2026

Abstract Paper Code

Deep learning models for medical image segmentation lack mechanisms to assess their own reliability, leading to two critical failures: they provide no uncertainty estimates to distinguish confident predictions from error-prone ones, and often produce anatomically implausible segmentations or incorrect connectivity that violate known structural constraints. We observe that uncertainty and topology are intrinsically linked and anatomically complex regions naturally exhibit higher prediction uncertainty, while uncertain predictions require stronger enforcement of structural constraints. Building on this insight, we propose TUNE++, a unified framework that jointly learns segmentation, uncertainty quantification, and topology preservation through a novel Topology-Uncertainty aware Paired Attention (TUPA) mechanism. Our method decomposes uncertainty into aleatoric and epistemic components while simultaneously enforcing anatomical correctness through persistent homology-based constraints. A key innovation is our topology-uncertainty alignment loss that minimizes the discrepancy between predicted total uncertainty and a topological complexity score computed from organ boundaries, multi-organ junction counts, and critical points extracted from persistence diagrams, teaching the model to be uncertain precisely where anatomical structure is geometrically complex. Our empirical results demonstrate that joint modeling of TUNE++ produced enhanced segmentation accuracy, well-calibrated uncertainty estimates that successfully identify errors, substantial reduction in topological violations, and learned confidence that correlates strongly with anatomical complexity.

HISTO-UNet: Histopathology Image Segmentation Using Topology-Aware UNet with Dual Uncertainty Quantification

Ashim Dhor, Emily Das, Rasel Mondal, Sweta Azad, Vaishali Walke, Shakti Kumar Yadav, Abhirup Banerjee, Tanmay Basu

IEEE International Symposium on Biomedical Imaging (ISBI) 2026

Abstract Paper Code

Histopathology image segmentation requires not only accurate pixel-wise predictions but also the preservation of topological structures and the quantification of prediction uncertainty for better diagnostics. We present HISTO-UNet, a novel framework that addresses these challenges by integrating topology-preserving constraints with a dual uncertainty quantification system for robust histopathology image segmentation. Our approach employs a multi-task objective, combining medial axis and marker-controlled topology losses with a Bayesian deep learning methodology to simultaneously capture both aleatoric and epistemic uncertainty. Through extensive evaluation on three public datasets, we found HISTO-UNet consistently outperforms standard baselines in both segmentation accuracy and uncertainty calibration. Our experiments validate the contribution of each component, revealing that the integration of topology-awareness and dual uncertainty quantification yields significant improvements in both segmentation performance and model reliability.

Topo-GraT: Learning to Grow with Causal Graph Transformers

Ashim Dhor, Smily Bharadwaj

AAAI 2026 — Student Abstract Track

Abstract Paper Code

Automated cancer segmentation in Whole Slide Images (WSIs) has been dominated by a paradigm of static pattern recognition, where even advanced methods leveraging Transformers, Multiple Instance Learning, or topology-aware losses remain fundamentally descriptive and correlational. To address this limitation, we reframe WSI segmentation from a descriptive task to one of causal process modeling. We introduce Topo-GraT, a novel framework featuring a Causal Growth Field (CGF) to model tumor invasion dynamics and a Causal Flow Attention (CFA) mechanism that embeds this field as an architectural prior. This causal engine is integrated within an iterative graph refinement loop that uses segmentation uncertainty to dynamically focus computational resources on the most ambiguous tissue regions. Our comprehensive experiments on multiple WSI datasets demonstrate that Topo-GraT establishes a new state-of-the-art, significantly outperforming existing methods and reducing the 95% Hausdorff Distance, a key boundary metric, by over 15%. Crucially, our framework yields the CGF as a rich, interpretable output whose structure correlates with tumor aggressiveness, positioning it as a novel biomarker for downstream prognostic tasks.

An Uncertainty-Aware Deep Learning Model for Gland Segmentation in Prostate Histopathology Images

Ashim Dhor, Emily Das

AAAI 2026 — Empowering Global South AI Community Activity

Abstract

Deep learning models for histopathology gland segmentation achieve high accuracy but lack confidence estimates, leading to silent failures in ambiguous regions and limited clinical trust. We propose an uncertainty-aware UNet++ framework that performs accurate gland segmentation and generates pixel-wise uncertainty maps to identify challenging regions. The framework explicitly models aleatoric uncertainty through Gaussian noise injection in logit space, combining cross-entropy, Dice, and uncertainty loss functions to optimize both segmentation accuracy and calibrated uncertainty estimation. Our framework achieves superior performance and low ECE, outperforming baseline models and existing uncertainty-aware methods including PHiSeg. This work provides a practical foundation for trustworthy AI in clinical pathology by offering transparent, confidence-calibrated predictions that help pathologists assess reliability and flag annotation issues.

2024

Financial Fraud Detection Using Quantum Graph Neural Networks

N. Innan, A. Sawaika, A. Dhor, S. Dutta, S. Thota, H. Gokal, N. Patel, M. A. Khan, I. Theodonis, M. Bennai

Quantum Machine Intelligence, 2024

Abstract Paper

Financial fraud detection is essential for preventing significant financial losses and maintaining the reputation of financial institutions. In this paper, we propose a novel approach for detecting financial fraud using Quantum Graph Neural Networks (QGNNs). QGNNs are a type of neural network that can process graph-structured data and leverage the power of Quantum Computing to perform computations more efficiently than classical neural networks. Our approach uses Variational Quantum Circuits (VQC) to enhance the performance of the QGNN. The results of our experiments showed that QGNNs achieved an AUC of 0.85, which outperformed classical GNNs. Our research highlights the potential of QGNNs and suggests that QGNNs are a promising new approach for improving financial fraud detection.

media highlights

Our paper HISTO-UNet: Histopathology Image Segmentation Using Topology-Aware UNet with Dual Uncertainty Quantification has been highlighted by The Times of India, Dainik Bhaskar and Research Matters. Heartfelt thanks to the entire team ; students, clinicians and collaborators at IISER Bhopal, AIIMS Bhopal, JNCHRC Bhopal and the University of Oxford; for making this possible. Some more works are going on…

work in progress

Large-Scale Whole Slide Image Dataset for Head & Neck and Multi-Organ Cancer

Ashim Dhor, Tanmay Basu, and clinical collaborators

In collaboration with AIIMS Bhopal & JNCHRC Bhopal

Ongoing

Building a large-scale, clinician-annotated dataset of whole slide images (WSIs) for cancer research. The primary focus is on Head & Neck cancers — including squamous cell carcinoma, adenoid cystic carcinoma, mucoepidermoid carcinoma, and verrucous carcinoma — with ongoing expansion to Breast cancer, Lung cancer (adenocarcinoma & squamous cell), Ovarian cancer, Prostate cancer, and Liver cancer (hepatocellular carcinoma). The dataset is being developed in close collaboration with pathologists at AIIMS Bhopal and Jawaharlal Nehru Cancer Hospital & Research Centre to ensure high-quality annotations with clinical-grade ground truth.

Collaborative Multi-Agent Reasoning for Whole-Slide Histopathology Analysis

Ashim Dhor, et al.

Manuscript in preparation

Ongoing

Developing a concept-specialized multi-agent framework for whole-slide histopathology that moves beyond monolithic foundation models. We apply non-negative matrix factorization on a multi-teacher ensemble to discover six interpretable morphological specializations — nuclear features, tissue architecture, stromal patterns, inflammatory response, necrosis, and normal tissue — and distill them into lightweight student agents. A learned query-aware router dynamically activates task-relevant agents per case, and selected agents engage in a Chain-of-Discussion protocol, iteratively exchanging structured evidence, refining predictions through attention-weighted integration, and explicitly flagging uncertainty when consensus cannot be reached — mirroring how pathologists deliberate on challenging cases.

Deliberative Multi-Agent Framework for Hypothesis-Driven Cancer Diagnosis

Ashim Dhor, et al.

Manuscript in preparation

Ongoing

Developing a multi-agent deliberative framework that decomposes WSI-based cancer diagnosis into four specialized agents mirroring expert clinical reasoning. A Perception Agent reduces gigapixel slides to salient regions via learned filtering. A Morphology Agent interprets cellular features through Chain-of-Thought reasoning with calibrated confidence. A Knowledge Agent grounds interpretations in authoritative medical literature via retrieval-augmented generation. A Deliberation Agent performs final diagnosis through Tree-of-Thought reasoning — maintaining competing hypotheses, adaptively selecting evidence to maximize information gain, and revising beliefs via Bayesian updating until reaching a confident diagnosis, differential, or principled abstention.

Uncertainty-Guided Adaptive Refinement for Brain Lesion Segmentation

Ashim Dhor, et al.

Manuscript in preparation

Ongoing

Developing a stroke lesion segmentation framework that explicitly models predictive uncertainty and leverages it as an active mechanism for adaptive refinement. SURE-Net integrates uncertainty-aware coarse segmentation, selective patch-based refinement guided by uncertainty maps, and topology-constrained boundary optimization into a unified pipeline. Rather than treating uncertainty as a passive byproduct, the framework uses it to allocate computational resources to ambiguous regions and enforce anatomical plausibility precisely where it is most needed — providing interpretable confidence estimates alongside segmentation outputs for clinical workflows.