Predicting the return of oral cancer.

For a patient treated for oral squamous cell carcinoma, the hardest question isn't the diagnosis — it's what comes after. Will the cancer come back, and how soon? The answer decides how closely they need to be watched.

§1The problem

Oral squamous cell carcinoma is among the most common cancers worldwide, and recurrence is common and dangerous. Clinicians estimate recurrence risk from experience and staging, but the signal is scattered across two very different kinds of data: structured clinico-pathological charts — stage, margins, node status, demographics — and unstructured histopathology images of the tumour tissue. Neither alone tells the whole story.

§2The approach

I built a multimodal machine-learning tool that predicts recurrence-free survival by fusing both data streams. Tabular clinico-pathological features are encoded alongside deep features extracted from histopathological images, and the combined representation feeds a survival model that estimates when — not just whether — recurrence is likely.

The imaging branch grew out of a companion study benchmarking deep learning architectures for classifying oral lesions from histopathology — work that later fed into an attention-guided, bias-mitigated classifier published in Nature Portfolio's Scientific Reports.

§3Context & collaborators

The project was carried out at the Indian Institute of Technology Jodhpur (2020–2022), advised by Dr. Dip Sankar Banerjee and Dr. Dharma Ram Poonia, bridging computer science and surgical oncology. It shaped the through-line of my research ever since: machine learning that clinicians can actually act on, applied to the diseases of the mouth.

§4The through-line

This is where the story starts. Predicting recurrence taught me that a prediction without a trustworthy confidence is only half an answer — the question that now drives my doctoral work on uncertainty quantification of large language models.