15 open-source projects · Python & deep learning

Medical AI that ships
with the paperwork to prove it

I build deep-learning systems for clinical imaging and the FDA/ISO validation tooling that gets them audit-ready — from chest X-ray and melanoma classifiers with explainability, to SaMD validation kits, pFMEA generators, and a public CDC disease-surveillance dashboard.

Explore the work View GitHub profile
15
Public repositories
6
Medical imaging models
5
FDA / ISO quality tools
★
Total GitHub stars
Py
Primary language
Loading live data from GitHub…
Medical Imaging, Signals & Deep Learning
6 projects · clinical imaging & physiological-signal models
thoravisPython
Multi-label chest X-ray classifier. ViT-B/16 backbone, CLAHE preprocessing, Grad-CAM heatmaps over the 15-class NIH ChestX-ray14 dataset.
PyTorchViTGrad-CAM
View repo →
Melanoma-Skin-ClassifierPython
ResNet-18 melanoma classifier on HAM10000 with Grad-CAM heatmaps, lesion bounding boxes, and per-prediction confidence percentages.
ResNet-18HAM10000XAI
View repo →
xray-gradcamPython
Grad-CAM explainability tooling for chest X-ray pneumonia detection — visualizing the regions driving each model decision.
Grad-CAMPneumonia
View repo →
polyp_segmentationPython
LightUNet experimentation for colonoscopy polyp segmentation — a lightweight architecture for pixel-level detection.
LightUNetSegmentation
View repo →
Reti-ves-segPython
Retinal blood-vessel segmentation on DRIVE. 2-channel CLAHE + Frangi U-Net (F1 0.833 inside FOV), served via a FastAPI overlay endpoint with label-free quality scores. Live project page at lsaiko.github.io/Reti-ves-seg.
U-NetCLAHE+FrangiFastAPI
View repo →
Opti-TracTPython
Real-time laparoscopic surgical instrument detection and tracking using YOLOv8s paired with ByteTracker.
YOLOv8sByteTrackReal-time
View repo →
Arry-fibulatorPython
Patient-specific ECG arrhythmia detection for pre-hospital monitoring. Calibrate-then-detect anomaly scoring with episode debouncing and rhythm-pattern features (RR sample-entropy, alternation) — 100% per-patient episode detection on MIT-BIH DS1/DS2. Live project page at lsaiko.github.io/Arry-fibulator.
ECGAnomaly detectionMIT-BIH
View repo →
Bone-RPython
Bone-fracture detection, localization, and characterization on X-rays. YOLOv8m trained on a harmonized FracAtlas + GRAZPEDWRI-DX merge, with FracAtlas bounding-box annotations and MURA image-level labels for robustness. Research/educational only — not a medical diagnosis.
YOLOv8mFracAtlas+MURADetection
View repo →

About this work

One engineer, two disciplines

The portfolio sits at an unusual intersection: the deep-learning skill to build clinical models, plus the regulatory fluency to validate them for real-world medical-device use.

Clinical Imaging AI

Classification, segmentation, and tracking across radiology, dermatology, endoscopy, and surgery — consistently paired with explainability (Grad-CAM, bounding boxes, confidence scores) rather than black-box outputs.

Regulatory & Quality

Tooling that targets the standards real medical-device teams answer to: FDA 21 CFR 820, ISO 13485, ISO 14971, IEC 62304, AIAG MSA, and AIAG-VDA pFMEA — turning compliance artifacts into generated, repeatable outputs.

Data & Public Health

Live, source-attributed data products — most prominently a CDC-backed disease surveillance dashboard with AI summaries — built to be free, transparent, and usable without an account.

Note on intended use: these projects are research, educational, and tooling artifacts. The clinical models are not cleared diagnostic devices, and the compliance generators produce preliminary drafts that require qualified human review before any regulatory submission.