Projects

Radiomics risk stratification in head and neck cancer

ESTRO 2026 poster highlight

• Built a 10-feature signature (2 clinical, 8 radiomic) to predict locoregional recurrence.
• Compared 5 metaheuristic algorithms, including a genetic algorithm, over 1,000+ candidate features.
• AUC 0.82 (95% CI 0.62–0.95) in 327 patients treated with primary radiation, with SHAP explanations for each prediction.

scikit-learn, NumPy, Pandas, SciPy

Reproducibility study of a CNN prognosis model

Under submission, Scientific Reports

• Re-ran the authors’ released code, data and Docker container. AUCs differed from the published ones by −0.22 to +0.10, and fixing the code issues took almost seven weeks.
• External validation on a separate cohort (n=102) gave an AUC of 0.49–0.59, close to chance.
• Scored the paper against CLAIM and TRIPOD+AI: high checklist compliance did not mean it was reproducible.

CT tumour segmentation with 3D CNNs

With NIT Karnataka, Surathkal

• Curated 167 CT cases (137 MAASTRO, 30 CMC Vellore) and trained a 3D nnU-Net for gross tumour volume. Dice was 0.62 on MAASTRO, 0.63 on CMC and 0.65 combined.
• Co-author of AGSR-Net, which lifts Dice from 0.70 to 0.73 and cuts HD95 from 33.92 to 21.22 mm across 896 patients.

nnU-Net, 3D Slicer, ITK-SNAP

Prospective imaging study, 2,000+ patients

DBT/Wellcome Trust India Alliance

• Built the DICOM retrieval pipeline on Orthanc PACS, then automated segmentation and PyRadiomics extraction for the whole cohort.
• Set up AWS S3 backups with automatic error recovery for a five-year, multisite radiotherapy study.

Python, Orthanc, DICOM/NIfTI, PyRadiomics, Bash

CHAVI, India’s first national oncology imaging biobank

• Contributed 304+ de-identified cases with validated clinical and imaging metadata, anonymised with XNAT and Orthanc.
• Worked with Tata Memorial Centre and IIT Kharagpur to harmonise data across institutions.

MGAttCNNMD: multimodal survival prediction

M.Tech thesis, Aligarh Muslim University, advised by Prof. Rashid Ali

• Three gated-attention CNNs (clinical, copy number, gene expression) fused by a stacked random forest.
• 91.2% accuracy for breast cancer survival on METABRIC, ahead of every single-modality baseline.