Akhil Veluru: New Generative AI Approach Advances Medical Vision and Image Analysis

Healthcare imaging AI solutions are being designed for a growing range of tasks, rather than to work on a single application moving to more generalised systems. The complex transformation of the domain from human-centered diagnoses to computerized diagnostics has escalated the need for flexible approaches in the interpretation, transformation and analysis of multi-modal medical images. For instance, a professional data and analytics professional is taking his technical journey further into medical vision research where he explores how generative AI and in-context learning could lead to more unified and scalable approaches to medical image analysis.

He has worked in various technologies, data rich and or involves analytics problems. Currently, he is working at Capital One as Data Warehouse Developer. In addition, he works on Snowflake in

financial services. Moreover, he creates as well as retains quality-assurance metrics. He also performs data refreshes, optimizes SQL workflows, and develops Tableau dashboard for stakeholders. He implemented over 10 QA metrics, resulting in a reported improvement of operational efficiency by 20%. Moreover, he obtained 99% availability for the primary reporting processes and improved refresh rates by 30%, among many other achievements. This shows a commitment to data and technology contextually and with structured data to improve visibility and decision-making.

The earlier positions in tech and financial analytics made that focus possible. While working at Wealus Technology Solutions as a Data Consultant, he has worked with AWS, Snowflake, SQL, SAS, Python, and Tableau to analyze driver revenue and profit, forecast sales performance, automate analytical tasks and support business-related decision making.

During his stint at Ernst & Young LLP, he was involved in tax reporting, ETL processes, analysis in SQL, reporting in Tableau, etc. Previously, he utilized the Excel algorithms and the SAP enterprise resource planning system for audit legal agreements for financial variances, along with supply-chain analytics at Hewlett-Packard Inc. The skills and expertise gained from this work has built a foundation in data quality.

Through academic preparation, this foundation was enhanced. Upon obtaining a bachelor of Business Management from Bangalore University, he completed a Master of Science in Business Analytics at The University of Texas at Dallas. His technical capabilities are SQL, Python, Snowflake, SQL Server, MySQL, Oracle Database, AWS, Azure, Tableau and Power BI. The candidate also holds AWS Certified Data Engineer Associate certification and CFA Level I which is a blend of data engineering, data analytics, and financial-domain related knowledge.

The transition towards data-driven problem-solving is now extending into medical artificial intelligence with the project “Unified Image-to-Image Generation for Diverse Medical Vision Tasks”. This paper introduces a framework called Medical Vision Generalist (MVG) that treats a number of medical vision problems as conditional image generation tasks. MVG integrates segmentation,

cross-modal synthesis, inpainting, and denoising into a single image-to-image generation framework across CT, MRI, X-ray, and micro-ultrasound, instead of using a separate architecture for each application.

At the heart of the project is in-context learning. The MVG employs visual prompts for task specification so that retraining the system for every task and having task-specific output heads is not necessary. It comes with a Vision Transformer backbone that combines masked image modelling and auto-regressive learning to balance understanding of local details and wider spatial relationships. This design overcomes a major hurdle in medical AI, which is that a specialist model can generally perform very well for a single task. However, different training with a different architecture is required, when the task or modality or dataset changes.

The study also prioritizes a wide scale evaluation instead of one dataset. A benchmark comprising 13 publicly available datasets corresponding to four anatomical regions, namely abdomen, pelvis, brain and chest, and four imaging modalities was developed. When combined, these datasets are rich with nearly 2.5 million training images. The assessment embraces segmentation and cross-modal synthesis, inpainting, denoising, and exploratory lesion detection, thus providing a varied ground for assessing generalisation.

The results that we report show the potential. MVG obtains an average mIoU of 0.79 across different segmentation tasks and outperforms a number of generalist baselines. For cross-modal synthesis, the results show a 0.002 reduction in MAE, a 0.69 increase in PSNR, and a 0.009 improvement in SSIM against the best generalist baseline reported in their work. The framework depicts a data-efficient adaptation of unseen Medical Segmentation Decathlon tasks. It achieves mIoU of 0.84 on MSD-Liver with a prompt and improves further on selected tasks with one-shot fine-tuning.

The importance of this work goes beyond the single model. The research direction of the project will aid in reducing task specialization and increasing transferability of AI systems across datasets and imaging conditions by exploring common frameworks for multiple medical vision requirements. The analysis of lesion localization via DeepLesion dataset provides additional insights on how visual prompting can be potentially helpful in different annotation formats.

To Akhil Veluru, this research is a natural progression for someone who has built a career transforming complex data into a structured actionable system. His journey from enterprise analytics and data warehousing to generative medical vision reflects an expanding technical scope linking business analytics to new AI research.

The MVG project places that experience within a health-care context, contributing to an ongoing effort at developing image-analysis systems for the medical domain that are adaptable, context-aware, and capable of operating in different clinical imaging settings.