Bishesh Khanalविशेष खनाल
Scientist · Research

From medical images to how work gets done

I started by modelling how the brain changes in Alzheimer's disease, moved to machine learning for fetal ultrasound, and then built a research program in Nepal on AI for health where specialists are scarce. My focus now is broader: how work itself is reorganized when any step can be held by a person or by an AI agent, and how people are skilled for that change.

70+papers & preprints
PhDINRIA, highest distinction
TOGAIthe group I lead at NAAMII: Transforming Global Health with AI
The arc

One question, asked at widening scales

2012–2016
INRIA, France

Modelling the brain

One of the first biophysical models of longitudinal brain deformation from atrophy in Alzheimer's disease (NeuroImage 2016), and a simulator of realistic longitudinal MRIs with known volume changes. PhD with Nicholas Ayache and Xavier Pennec.

2017–2019
King's College London · Imperial College London

Machine learning for fetal ultrasound

On the iFIND project: standard-plane detection in 3D fetal ultrasound, tracking in freehand 4D ultrasound without external trackers, ultrasound–MRI registration and multi-view compounding. Worked with Julia Schnabel, Bernhard Kainz and Daniel Rueckert.

2019–
NAAMII, Nepal

Health AI for low-resource settings

AI-powered task-shifting: letting nurses and frontline workers do what used to need a specialist. Blind-sweep obstetric ultrasound that non-experts can perform, smartphone microscopy for parasites, AI-assisted cervical screening, and 3D bone shape from two inexpensive X-rays (NeurIPS 2023). Many of these are among the first papers from Nepal in their fields.

2022–
NAAMII

Foundation models, efficiently

Making large vision-language models useful on small, noisy medical datasets and modest hardware: prompt tuning, lightweight adapters, and robustness to label noise.

2021–
International consortia

Trustworthy AI & policy

Co-author of the FUTURE-AI international consensus guideline (BMJ 2025) and of a framework for just and equitable health AI in South Asia (npj Digital Medicine 2025). Member of the committee that drafted Nepal's national AI policy.

Now
NAAMII · Tangible

Reorganizing work

Every organization is a set of processes: flows that turn data into changes in its records, through actors who may be people or AI agents. My current research follows one chain from start to end. Personalized learning on problems that are real in the organization changes how its processes run, and those reorganized processes then have to be built. In health, I call this implementation science for AI and health. The first paper is The AI Pyramid.

  1. Personalized, scalable learning
  2. Skilling on real problems
  3. Capability building
  4. Reorganized processes
  5. Built and deployed
Publications by theme

69 papers, 12 threads

1 paper

Reorganizing work

Workforce capability, skilling and the reorganization of work when a step can be held by a person or an AI agent.

17 papers

Fetal & obstetric ultrasound

From 3D fetal imaging in London to AI-guided blind-sweep ultrasound that non-specialists can perform.

6 papers

Frugal diagnostics

Point-of-care diagnostics on commodity hardware — smartphones, paper assays, ECG and telehealth screening.

5 papers

X-ray & 2D→3D

X-ray analysis and biplanar X-ray to 3D bone reconstruction where CT is unavailable.

7 papers

Foundation models, efficiently

Adapting vision-language and foundation models with prompts, adapters and parameter-efficient tuning.

10 papers

Learning from imperfect data

Label noise, weak and semi-supervision, active learning and uncertainty.

6 papers

Trustworthy AI & policy

Trustworthy, equitable and deployable AI — FUTURE-AI, South Asia, generative AI in LMICs.

4 papers

Nepali & low-resource language

NLP and LLM evaluation for Nepali and Devanagari-script languages, especially for health.

16 papers

Modelling & geometry

Biophysical models of brain atrophy, epidemiological models, shape and pose geometry.

12 papers

Medical image analysis

Segmentation and reconstruction methods across imaging modalities.

4 papers

Proceedings

Co-edited MICCAI workshop proceedings — ASMUS, FAIR, DART, DeCaF.

3 papers

Early computer vision

Where it started — vision-based robots, skin and road-sign detection.

Selected work

A few papers to start with

  1. Automated quality assessment of blind sweep obstetric ultrasound for improved diagnosis
    Bhandari P, Poudel K, Luitel N, Acharya B, Ghimire A, Wellman T, Koepsell K, Regmi PR, Khanal B
    Medical Imaging with Deep Learning (MIDL). Proc Mach Learn Res, 2026;315:2987-2997
  2. The AI pyramid: a conceptual framework for workforce capability in the age of AI
    Khatri A, Khanal B
    arXiv preprint
  3. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare
    Lekadir K, Frangi AF, Porras AR, Glocker B, Cintas C, Langlotz CP, Weicken E, Asselbergs FW, Prior F, Collins GS, Kaissis G, Tsakou G, Buvat I, Kalpathy-Cramer J, Mongan J, Schnabel JA, Kushibar K, Riklund K, Marias K, Amugongo LM, …, FUTURE-AI Consortium (incl. Khanal B)
    BMJ, 2025;388:e081554
  4. Transforming healthcare through just, equitable and quality driven artificial intelligence solutions in South Asia
    Adhikari S, Ahmed I, Bajracharya D, Khanal B, Solomon C, Jayaratne K, Mamum KAA, Talukder MSH, Shakya S, Manandhar S, Memon ZA, Chowdhury MH, Ul Islam I, Rakhshani NS, Khan MI
    NPJ Digit Med, 2025;8(1):139
  5. VLSM-adapter: finetuning vision-language segmentation efficiently with lightweight blocks
    Dhakal M, Adhikari R, Thapaliya S, Khanal B
    MICCAI. 2024:712-722. Springer
  6. Benchmarking encoder-decoder architectures for biplanar X-ray to 3D bone shape reconstruction
    Shakya M, Khanal B
    Adv Neural Inf Process Syst (NeurIPS), 2023;36:20469-20481
  7. Standard plane detection in 3D fetal ultrasound using an iterative transformation network
    Li Y, Khanal B, Hou B, Alansary A, Cerrolaza JJ, Sinclair M, Matthew J, Gupta C, Knight C, Kainz B, Rueckert D
    MICCAI. 2018:392-400. Springer
  8. A biophysical model of brain deformation to simulate and analyze longitudinal MRIs of patients with Alzheimer's disease
    Khanal B, Lorenzi M, Ayache N, Pennec X
    NeuroImage, 2016;134:35-52
All publications →
Projects

What the lab is building

AI for ultrasound

Making portable, low-cost ultrasound usable by non-experts: acquisition, quality control, gestational age and biometrics from blind sweeps.

Setting change

Can we know, before retraining, how a model will perform in a new hospital or country, and adapt it on the smallest budget?

Workforce & health AI

Agentic health-AI systems, and the training and workflow changes they need to actually work in a clinic.

Frugal diagnostics

Smartphone microscopy, paper-based assays and AI-assisted cervical screening, built for places without a lab.

Foundation models at the edge

Compressing multimodal foundation models so they are more accurate with fewer parameters, on devices that fit the setting.

CDiTH

As consultant advisor: endometriosis, surgical-video understanding, and evaluation of hospital AI.

Team & alumni

The TOGAI team, and where they went

Since 2019: about 20 research assistants at NAAMII and 2 at IIIT Hyderabad, 5 MSc theses and 20 interns. I also give visiting lectures at the University of Cambridge (MSc, AI in global health, 2024–) and gave them at Imperial College London (PhD students, AI in ultrasound imaging, 2018).

Current TOGAI team

Alumni, and where they are now

Service to the field

Co-chair of the ASMUS, FAIR and MIRASOL workshops at MICCAI, the MICCAI Young Scientist Publication Impact Award committee (2025), and reviewer for NeurIPS, CVPR, MICCAI, IEEE TMI and others. Service & policy →

Funders & partners

Gates Foundation · GE HealthCare · IDRC / Grand Challenges Canada · Lacuna Fund / Wellcome · AmplifyChange · The Asia Foundation · UNDP Nepal · NAST · ETH4D (ETH Zurich) · AIMIX ERC via the University of Barcelona · NVIDIA