Portfolio Home Research Contact
suchibratapatra2003@gmail.com GitHub LinkedIn
Research & Publications — Suchibrata Patra
4 Manuscripts • Medical Imaging • Climate Science • Computational Cognition
Portfolio › Research & Publications

Manuscripts & Research Output

Four manuscripts in preparation spanning neuro-oncology imaging, spatiotemporal deep learning, and computational aesthetics. Primary focus on glioblastoma characterization using radiomics and pharmacokinetic modelling at TCG CREST in collaboration with the University of Pennsylvania.

Medical Imaging Neuro-Oncology & Radiomics 2 papers
Neuro-Oncology
Manuscript in Preparation — Second Author

Differentiating True Progression from Pseudoprogression in Glioblastomas using DCE-MRI

A critical unresolved challenge in post-treatment glioblastoma management is the differentiation of true tumour progression from pseudoprogression — a treatment-induced inflammatory response that produces imaging changes indistinguishable from disease recurrence on conventional MRI. This work applies dynamic contrast-enhanced MRI with voxel-wise parsimonious Tofts pharmacokinetic modelling to extract temporal biomarkers including Ktrans, Ve, and Vp for multiparametric habitat-based classification. Automated nnU-Net tumour segmentation pipelines are integrated with high-dimensional radiomic feature extraction and ensemble-based machine learning for non-invasive treatment response prediction.

Contributions
  • Developed pharmacokinetic models for temporal biomarker extraction from DCE-MRI sequences using voxel-wise Tofts modelling.
  • Built complete MRI preprocessing, bias-field correction, co-registration, and tumour segmentation workflows using SimpleITK and nnU-Net.
  • Engineered multiparametric habitat-imaging framework combining PK-parameter maps with radiomic features from PyRadiomics.
  • Designed ensemble-based classification pipeline with LightGBM, Random Forest, and SHAP-based interpretability.
DCE-MRI PyRadiomics nnU-Net PK Modelling Habitat Imaging LightGBM TCG CREST × UPenn — In preparation
Status Manuscript in preparation — Second Author • TCG CREST × University of Pennsylvania
Machine Learning
Significant Contribution

Glioblastoma Radiomics Classification using Machine Learning

High-dimensional radiomic feature extraction across multimodal MRI sequences (T1ce, T2-FLAIR, T1, ADC) for non-invasive glioblastoma tissue characterization. The study engineers 150+ radiomic features per modality using PyRadiomics, applies LASSO-based feature selection under 5-fold cross-validation, and constructs an ensemble classifier combining LightGBM and Random Forest. SHAP values provide patient-level feature attribution aligned with clinical interpretability requirements. The final model achieves AUC 0.89 with strong precision-recall characteristics on the held-out test set.

Contributions
  • Engineered 150+ radiomic features across multimodal MRI sequences using PyRadiomics (first-order, GLCM, GLRLM, shape features).
  • Achieved AUC 0.89 using LightGBM and Random Forest under rigorous 5-fold cross-validation with stratified splits.
  • Implemented SHAP-based global and local feature attribution framework mapping radiomic features to clinical outcomes.
  • Applied LASSO regularization for high-dimensional feature selection prior to classifier training.
LightGBM Random Forest SHAP AUC 0.89 150+ Features LASSO Significant Contribution — In preparation
Key Result AUC-ROC: 0.89 • 150+ radiomic features • LightGBM + Random Forest ensemble
Climate Science Spatiotemporal Deep Learning 1 paper
Climate & Deep Learning
Manuscript in Preparation — First & Corresponding Author

Forecasting Dry and Humid Heatwaves over India Using Physics-Guided Spatiotemporal Deep Learning

Heatwaves are among India’s most lethal meteorological hazards, yet systematic differentiation of dry versus humid heatwave events using data-driven approaches remains computationally understudied. This work designs a ConvLSTM–Transformer hybrid architecture for categorical heatwave forecasting using ERA5 reanalysis data (0.25-degree spatial resolution), incorporating atmospheric physics as inductive bias through thermodynamic and hydrological threshold-based labelling. XAI tools including SHAP and Grad-CAM evaluate learned atmospheric mechanisms and assess model interpretability against known physical processes governing heatwave dynamics over the Indian subcontinent.

Contributions
  • Designed ConvLSTM–Transformer hybrid architecture for spatiotemporal heatwave forecasting from ERA5 reanalysis data.
  • Developed physics-guided feature selection and categorical heatwave labelling based on thermodynamic and hydrological thresholds.
  • Applied SHAP and Grad-CAM explainability tools to evaluate learned atmospheric mechanisms and model interpretability.
  • Constructed preprocessing pipeline handling 0.25-degree ERA5 gridded reanalysis data over the Indian domain.
ConvLSTM Transformer ERA5 SHAP Grad-CAM Physics-Guided ML First & Corresponding Author — In preparation
Data ERA5, ECMWF Copernicus Climate Data Store • First & Corresponding Author
Computational Cognition Visual Representation & Perturbation 1 paper
Computational Aesthetics
Manuscript in Preparation — First & Corresponding Author

Fragility of Aesthetic Judgment vs. Stability of Emotional Representation under Visual Perturbation

A fundamental open question in computational aesthetics and visual representation learning concerns the differential robustness of neural encodings trained on distinct supervisory signals. This work systematically investigates the representational stability of CNN embeddings trained under aesthetic versus emotional supervision using controlled Gaussian noise perturbation tests. Embeddings are probed at multiple noise intensities to characterize the rate of representational collapse, revealing that emotional representations exhibit significantly greater stability than aesthetic representations under progressive visual degradation — with implications for the design of robust visual AI systems deployed in noisy real-world environments.

Contributions
  • Designed experimental framework investigating representational stability of CNN embeddings under aesthetic vs. emotional supervision.
  • Conducted Gaussian noise perturbation tests across multiple noise intensities to characterize representational collapse rates.
  • Demonstrated that emotional representations remain stable under visual degradation while aesthetic representations degrade sharply.
  • Derived implications for robust visual AI system design in noisy real-world deployment settings.
CNN Embeddings Perturbation Analysis Representation Learning Gaussian Noise Visual AI First & Corresponding Author — In preparation
Applied Projects Independent Research & Competition Work 3 projects
Competition
RedBus ML Hackathon — Nationwide Rank #43

Demand Forecasting for Intercity Bus Routes using Hybrid Time-Series Models

Built hybrid forecasting pipelines combining SARIMA, Prophet, and LSTM architectures for intercity bus demand prediction. Achieved 21% accuracy improvement over baseline ARIMA and secured Nationwide Rank #43 among 800+ competing teams.

SARIMAProphetLSTM+21% AccuracyRank #43 • 800+ Teams
Clinical ML
Independent Project

Cardiovascular Disease Risk Prediction via Logistic Regression with Clinical Feature Engineering

Clinical risk model using logistic regression with feature engineering, interaction terms, and comprehensive ROC-AUC evaluation for clinical decision support in cardiovascular risk stratification.

Logistic RegressionFeature EngineeringROC-AUC
ML Engineering
Independent Project

Customer Churn Prediction: SMOTE-Balanced Ensemble Pipeline with False-Negative Minimization

Full ML pipeline using SMOTE for class imbalance correction, advanced feature engineering, and gradient-boosted ensemble models. Achieved 18% reduction in false negatives relative to baseline logistic regression.

SMOTEEnsemble Models−18% False Negatives