Suchibrata Patra

Between a physician's judgment and a patient's outcome lies a single, largely invisible variable i.e the quality of the information connecting them. This understanding is the primary motivation behind my research. Growing up, I observed my mother working in public health administration under the Government of India, which helped me realise that both clinical and policy decisions are only as reliable as the data they are based on. This perspective has guided my academic interests. I was never drawn to clinical medicine itself. My instinct has always been toward numbers, patterns, and the logic underneath complex systems. My engagement with this field is therefore not incidental, but grounded in an early and concrete understanding of its purpose. That conviction found its direction during my M.Sc. in Data Science at St. Xavier's College (Autonomous), Kolkata, where I worked concurrently as an AI Research Intern at TCG CREST under the supervision of Dr. Sourav Bhaduri. My master's thesis addressed a clinically significant problem in neuro-oncology: the reliable discrimination of treatment response, a treatment-induced inflammatory response that standard contrast enhanced MRI cannot differentiate with adequate diagnostic confidence. Misclassification in this context carries direct clinical cost. A patient may be withdrawn from an radio-chemotherapy treatment, or subjected to escalated intervention on the basis of a misinterpreted imaging finding. To address this, I developed a computational pipeline that extracts quantitative radiomic features, encompassing intensity, texture, and shape descriptors, from dynamic MRI sequences using a radiomics-based approach, and integrates them with molecular markers. The underlying hypothesis was that a more biologically grounded representation of tumour state would yield more reliable diagnostic decisions. The results supported this. This finding is currently being prepared for submission to a peer-reviewed journal. That work, however, also delineated the boundaries of what the current approach can offer. The dataset was drawn from a single centre and a single timepoint. This was appropriate for a master's thesis, but it is insufficient for models intended for broader clinical deployment. Single-site models routinely fail to generalise across institutions, imaging protocols, and patient cohorts. That ceiling on generalisation is, ultimately, a ceiling on clinical utility. My proposed research aims to address this systematically. I intend to build and validate diagnostic frameworks on larger, multi-institutional medical imaging datasets, with the explicit objective of developing tools that perform reliably across heterogeneous clinical environments. A related limitation is that my prior work treats each imaging acquisition as an independent observation. This is a methodological simplification that does not reflect biological reality. Disease progression is a temporally structured process. A model that cannot account for how a patient's condition evolves between timepoints is, by design, working with an incomplete picture. A model that processes sequential imaging data is fundamentally better positioned to separate treatment response. I intend to develop temporally aware frameworks that encode this longitudinal structure, learning not from isolated snapshots but from how a patient's condition evolves across timepoints. The clinical benefit is direct. Fewer patients misclassified, fewer unnecessary interventions, and treatment decisions grounded in a more accurate and complete reading of disease trajectory. Integrating molecular and clinical markers alongside imaging features represents a natural extension of this work. My prior findings suggest that richer multimodal representations of disease state yield models that are both more accurate and more generalisable. These directions, taken together, follow a coherent internal logic. Each addresses a specific inadequacy of the current paradigm. Each moves in the same direction: from models optimised for the controlled conditions of a single study, toward frameworks that remain reliable across the heterogeneity of real clinical environments. That transition, from methodological proof of concept to genuine clinical utility, is where the field's most consequential work currently lies. I am applying to the Ph.D. programme at the Institute for Advancing Intelligence (IAI), TCG CREST, under the supervision of Dr. Sourav Bhaduri, for reasons that go beyond academic continuity. My internship at TCG CREST shaped how I think about the relationship between methodological rigour and clinical consequence. Working within this lab gave me more than technical training. It gave me a clear understanding of what good translational research looks like in practice, and where the hardest problems still lie. I am familiar with the lab's research priorities, its working culture, and the kind of questions it considers worth pursuing. That familiarity matters. I am not applying to learn a field from scratch. I am applying with a defined set of questions, prior experience engaging with them, and a clear account of what remains to be done. IAI's explicit focus on AI and Machine Learning , including Computer Vision, Image Analysis, and Medical Imaging, combined with its SOTA GPU infrastructure and its culture of rigorous, thesis-based doctoral research, makes it uniquely suited to the programme I propose. I believe this institution and this mentorship offer the environment best suited to pursuing it well.

• 2024
  2nd Runner-up — PharmaQuant Hackathon

  Top 3 of 800+ participants · Statistical analysis & quantitative pharmacology challenge
• 2024
  Nationwide Rank #43 — RedBus ML Competition

  Demand forecasting under distribution shift among 4000+ competing teams
• 2021
  District 2nd Rank — WBBSE Secondary Examination

I am open to PhD enquiries, research collaborations, and applied roles in medical imaging, computational oncology, or statistical data science. I respond to all substantive academic correspondence.