Nima Tajbakhsh

Timeline

1. 2016

   Polyp detection in colonoscopy

   In 2016, he publishes a computer-aided detection system for polyps in colonoscopy videos in IEEE Transactions on Medical Imaging, written with Suryakanth R. Gurudu and Jianming Liang. The hybrid context-shape system reaches 88.0% sensitivity on the CVC-ColonDB images at 0.1 false positives per frame.

2. 2016

   Fine-tuning or full training

   In 2016, he co-authors an IEEE Transactions on Medical Imaging study that asks whether pretrained convolutional networks with sufficient fine-tuning can replace training from scratch in medical imaging. Across four applications in radiology, cardiology and gastroenterology, a pretrained network with adequate fine-tuning outperforms or, at worst, matches a network trained from scratch.

3. 2018

   UNet++ for segmentation

   In 2018, he co-authors UNet++ with Zongwei Zhou, Md Mahfuzur Rahman Siddiquee and Jianming Liang, a deeply supervised encoder-decoder network with nested, dense skip pathways. Accepted at the 4th Deep Learning in Medical Image Analysis workshop, it gains an average of 3.9 IoU points over U-Net with deep supervision.

4. 2019

   Segmentation with imperfect datasets

   In August 2019, he and five co-authors post Embracing Imperfect Datasets, a review of deep learning solutions for medical image segmentation, accepted for publication in Medical Image Analysis. The review covers scarce annotations and weak annotations such as sparse, noisy and image-level labels.

5. 2021

   A guest editorial on annotation

   In 2021, he writes the IEEE Transactions on Medical Imaging guest editorial Annotation-Efficient Deep Learning: The Holy Grail of Medical Imaging with Holger Roth, Demetri Terzopoulos and Jianming Liang.

6. April 2024

   DRaFT+ in NVIDIA NeMo

   In April 2024, he writes an NVIDIA Technical Blog post with Ali Taghibakhshi, Sahil Jain, Gerald Shen and Arash Vahdat on DRaFT+, now part of NVIDIA NeMo. The method adds a regularization term to direct reward fine-tuning of text-to-image diffusion models to enhance generation diversity, control reward over-optimization and prevent mode collapse.

7. 2025

   Video training and SSM pruning

   In March 2025, he is among the 29 authors of Training Video Foundation Models with NVIDIA NeMo, which presents an open-source pipeline for video dataset curation, multimodal data loading and parallelized video diffusion model training and inference. In April 2025, he co-authors Minitron-SSM, which compresses the Nemotron-H 8B Hybrid model down to 4B parameters with up to 40x fewer training tokens.

8. As of October 2026

   Manager within NVIDIA NeMo

   He is a deep learning algorithm manager who specializes in computer vision, LLM and multimodal GenAI models. Within NVIDIA NeMo, he leads the development and optimization of training and inference workflows for diverse GenAI models.