Mayur Mallya

University of British Columbia
3 previous affiliations
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Mayur Mallya focuses on medical image analysis, multimodal deep learning, and computational pathology. Early collaborative work explored the application of artificial intelligence to neuro-oncology imaging, evolving into primary methodological contributions that leverage cross-modal guidance to improve diagnostic accuracy. Most recently, Mallya's efforts address the automated prediction of treatment effectiveness using gigapixel-scale histopathological images in precision oncology.

Research Trajectory

Mayur Mallya’s research trajectory reflects a sustained interest in bridging the gap between advanced machine learning techniques and practical clinical workflows. The available evidence outlines an evolution from broad clinical applications of artificial intelligence in radiology to targeted methodological improvements in multimodal learning and computational pathology.

Initial collaborative work centered on the intersection of neuro-oncology and radiological imaging. Mallya contributed to Applying Artificial Intelligence to Glioma Imaging: Advances and Challenges, a 2019 review examining how AI could assist clinicians in managing primary brain tumors. This work highlighted the necessity of using computed tomography (CT) and magnetic resonance imaging (MRI) for surgical planning and tracking tumor progression, while also acknowledging the hurdles of deploying these tools in clinical settings. This early participation aligns with the broader research direction of Learning from Limited Data in Specialized Medical Domains, where the data bottleneck often complicates the deployment of robust AI models for specialized tasks like glioma detection.

By 2022, Mallya’s focus shifted toward core methodological challenges in medical computer vision, specifically addressing the trade-offs inherent in different imaging modalities. As first author of Deep Multimodal Guidance for Medical Image Classification, Mallya proposed a framework to reconcile the tension between clinical feasibility and diagnostic performance. In clinical practice, highly accurate imaging modalities are often expensive, invasive, or slow to acquire, whereas more accessible modalities may yield inferior diagnostic signals. Mallya’s approach leverages knowledge learned from a superior, less feasible modality during training to guide and steer the utilization of an inferior, more feasible modality during inference. This work represents a significant contribution to the area of Deep Learning Uncovers Latent Prognostic Signals in Medical Imaging, demonstrating how latent representations can be transferred across modalities to improve theranostic tasks without requiring the superior modality at the time of diagnosis.

More recently, Mallya has participated in large-scale collaborative efforts in precision oncology and digital pathology. In 2025, Mallya was a co-author on ATEC23 Challenge: Automated prediction of treatment effectiveness in ovarian cancer using histopathological images. This study details the outcomes of a MICCAI challenge focused on predicting treatment efficacy from whole-slide images (WSIs) and tissue microarrays (TMAs). Mallya's involvement in this initiative reflects an engagement with the complex computational demands of digital pathology, a field closely tied to the research theme of Scaling Deep Learning to Gigapixel Whole Slide Images, where the sheer scale of tissue scans requires specialized architectures to process effectively.

Current Directions

The most recent papers and themes in the evidence suggest a current focus on the computational challenges of digital pathology and precision oncology. Mallya’s recent collaborative work investigates how high-dimensional biological realities, captured in gigapixel-scale tissue scans, can be distilled into actionable prognostic signals for patient selection.

A key component of this ongoing work involves navigating the immense memory requirements of whole-slide images. This aligns with active engagement in the research direction of Attention Based Multiple Instance Learning, where massive images must be divided into manageable patches for neural network analysis. Furthermore, the extraction of meaningful patterns from these complex domains intersects with themes such as Global Context and Long Range Dependencies in Vision Transformers and the broader area of Biological Feature Discovery through Transformer Attention Alignment. Through these avenues, current efforts aim to move beyond black-box predictions, facilitating treatment planning by uncovering the biological rationale embedded within histopathological data.

Profile updated

See something wrong? Report a mistake