I'm an M.S. student in Engineering AI at Carnegie Mellon University Africa. I build vision–language and multimodal models for medical imaging and clinical decision support. My work focuses on making these models efficient enough to run on large 3D scans, and honest enough about their limits to be trusted in low-resource health settings.

Research interests

• Medical vision–language models for 3D imaging
• Instruction-conditioned and token-efficient architectures for volumetric data
• Grounded radiology report generation and multi-scale representation learning
• Interpretable and reliable medical AI
• Deployment of imaging AI in low-resource and African clinical settings

ICTC: Instruction-Conditioned Token Compression for 3D CT Vision–Language Models
A dual-stage, instruction-conditioned token compressor for 3D chest CT. A Perceiver stage distils each slice and a FiLM-conditioned stage aggregates slices into a fixed budget of visual tokens for a LoRA-adapted language model that writes the radiology report. On CT-RATE validation (3,039 scans): ROUGE-L 0.273, BLEU-4 0.128, RaTEScore 0.801. Summer 2026 research internship with Prof. Prasenjit Mitra.

FPN-Mamba: Linear-Time Bidirectional State Space Feature Pyramids for Small-Lesion Brain MRI Classification
Feature pyramid plus bidirectional Mamba for small brain-metastasis screening on UCSF-BMSR. It outperforms a self-attention baseline (InceptentionNet) on AUC (0.982 vs 0.897) and smallest-lesion recall (0.870 vs 0.788) across 5 seeds × 3 folds, and is 3.7× faster per step than a matched attention mixer at 512 px. NeurIPS 2026 Africa in AI workshop.

Where Do Malaria Detector False Positives Come From?
A controlled study of two YOLOv12s detectors on 8,000 thick-smear images from Uganda and Ghana, using seven spatial and statistical analyses. It finds that white blood cells are not a systematic source of false positives, which redirects mitigation toward staining variability. MICCAI 2026 MIRASOL workshop (oral).

Condition-Aware ICU Deterioration Risk
Diagnosis-conditioned early-warning models on MIMIC-IV, benchmarking six sequence architectures at 1–3 h horizons. The best model (BiLSTM + attention, ROC-AUC 0.76) is served as a Dockerized FastAPI + WebSocket streaming service, with limitations documented up front.

Multimodal Cancer Survival Prediction (in progress)
Generalizable survival modeling from TCGA whole-slide images, RNA and clinical data, building on SurvPath.

• Investigating White Blood Cells as a Source of False-Positive Malaria Parasite Detection in African Blood-Smear Images
  Samuel Akinwumi Adeniji, Goodness Chidera Obasi, Chris-Victor Ntwali, Aondona Moses Iorumbur, Confidence Raymond, Lowami Uwimana, Ahmed Tahiru Issah
  MICCAI 2026 Workshop: MIRASOL (oral presentation) · MICCAI · arXiv · PDF · OpenReview · Code

• FPN-Mamba: Linear-Time Bidirectional State Space Feature Pyramids for Small-Lesion Brain MRI Classification
  Simbiat Damilola Adetoro, Samuel Akinwumi Adeniji, Moise Busogi
  NeurIPS 2026 Workshop: Africa in AI · OpenReview · Code

Full list on Google Scholar · CV

PyTorch · TensorFlow · Vision Transformers · Mamba / state-space models · CNNs and Feature Pyramid Networks · vision–language models · LoRA / parameter-efficient fine-tuning · HuggingFace · 3D CT and MRI pipelines · CT-RATE · MIMIC-IV · MONAI-style preprocessing · OpenCV · YOLO · scikit-learn · Weights & Biases · Docker

Website · Google Scholar · LinkedIn · sadeniji@andrew.cmu.edu