Most medical imaging AI stops at the benchmark. I build the part that comes after: segmentation pipelines validated against the thresholds clinicians actually use, diagnostic models that explain their reasoning and quantify their own uncertainty, and networks compressed to run on the hardware a hospital already owns.

Three principles run through every system here:

• Models that meet clinical standards. Fetal head circumference within ISUOG's ±3mm. ASD detection sensitivity and specificity reported per acquisition site, not pooled. If it doesn't pass the clinical bar, it doesn't ship.
• Explainability built in, not bolted on. GradCAM++/LIME spatial heatmaps, MC-Dropout uncertainty, and a structured PDF report that tells the clinician what the model found — and where it is unsure.
• Deployment-aware from day one. 2× CNN compression with no accuracy penalty, 0.033ms edge inference, live HuggingFace Spaces deployments, and Model Cards with regulatory framing (FDA SaMD Class II, EU IVDR Class B).

14 projects · 3 deployed systems · 5 imaging modalities — full write-ups at tarunsadarla2606.github.io.

Manual HC calliper placement takes 2–4 minutes per antenatal scan with up to 7mm inter-observer variation — roughly 153 sonographer hours per year at a unit doing 20 scans/day. This system automates it.

Residual U-Net with boundary-weighted loss and deep supervision · 16-frame temporal self-attention for cine-loop sweeps · Pseudo-LDDM v2 motion synthesis · Hadlock gestational-age estimation · GradCAM++ and per-frame uncertainty maps · clinical PDF report. Third-trimester MAE (7.60mm, acoustic shadowing from the ossified skull) is documented explicitly in the Model Card as a known limitation.

PyTorch U-Net Temporal Attention GradCAM++ Structural Pruning Streamlit HC18

ASD diagnosis takes 18–24 months from referral in many health systems, while structural MRI is already routinely acquired. This system reads that existing data as a pre-assessment triage layer.

Raw NIfTI volumes · 4-metric quality gate · GradCAM + LIME dual explanation with cross-validation · MC-Dropout uncertainty over 30 stochastic passes · LLM-generated clinical PDF with site reliability indicators. Site sensitivity spans 88.5% (PITT) to 98.5% (UM_1) — scanner heterogeneity treated as a deployment constraint, not a confound. Confident false positives (σ ≈ 0.005) documented as the dangerous failure mode.

PyTorch nibabel GradCAM LIME MC-Dropout ReportLab Streamlit ABIDE-I

The 2023 B.Tech origin version (CNN vs ViT, 91% accuracy) is preserved separately — the gap between that model and this system is what the MS was built to close.

Directed study · CSCE 5934 · IEEE-format report · individual contribution

When two convolutional filters are redundant, standard structured pruning permanently discards the weaker one. Hybrid Crossover synthesises a replacement instead — least-squares regression fits a new filter to the element-wise max of both parents' activation maps, preserving the peak response of each. Integrated into a global ILR saliency pipeline with hard accuracy guard rails (≤2pp overall, ≤6pp per-class).

2× channel compression · +0.37% accuracy · 2× faster runtime · 2.05× best latency speedup

The counterintuitive part is the speed: synthesised filters clear the guard rails on the first attempt, while standard deletion under aggressive compression triggers rollback loops that dominate total runtime. Motivated directly by the fetal HC deployment constraint — same clinical accuracy, half the GPU memory.

PyTorch VGG-16 ILR Saliency CelebA CIFAR-10

Course collections: Machine Learning (RBF SVM, VGG-16 transfer learning at 0.9594 acc, Bayesian networks) · Fundamentals of AI (Dijkstra/A*, genetic scheduling, value-iteration RL) · AI in Wearables & Healthcare (FDA/SaMD frameworks, Seeed XIAO nRF52840 PPG+IMU, YOLOv8 mAP@50 0.913, MIMIC-IV) · Scientific Data Visualization (Tableau, D3.js, Power BI)

Medical Imaging & Vision — DICOM · NIfTI · Obstetric ultrasound · Structural MRI · Digital pathology · U-Net segmentation · CNNs · Vision Transformers · OpenCV

Explainability & Clinical Governance — GradCAM++ · LIME · MC-Dropout uncertainty · Model Cards · bias/subgroup audits · FDA SaMD Class II · EU IVDR Class B · IEC 62304

ML & Deep Learning — PyTorch · TensorFlow · Keras · scikit-learn · XGBoost · Temporal attention · Structural pruning · Knowledge distillation

Deployment & Cloud — HuggingFace Spaces · Streamlit · Django REST · React · Flask · FastAPI · Docker · TorchScript · AWS (S3, EMR)

Big Data & Tools — PySpark · Hadoop · Hive · SQL · Git · ReportLab

Languages — Python · MATLAB · SQL · C

MS Artificial Intelligence (Biomedical Concentration) — University of North Texas · GPA 4.00 · May 2026

Instructional Assistant, Fundamentals of AI (CSCE 5210) — University of North Texas · Aug 2025 – May 2026 Supported course instruction for Prof. Russel Pears across two semesters covering AI agent architectures, search methods, and optimization. Designed and graded assignments, held office hours, and gave one-on-one debugging support on search strategies and optimization algorithms.

B.Tech Computer Science & Engineering — Manipal Institute of Technology · Minor: Data Analytics · Dec 2023

Medical Imaging AI Engineer — radiology AI, segmentation pipelines, DICOM workflows ML Engineer, Healthcare — clinical decision support, model deployment, XAI AI Research Engineer — applied research across medical imaging, ultrasound, and pathology AI, with particular strength in oncology and cancer imaging Clinical AI Scientist — model validation, regulatory evidence, clinical integration