Haomin Chen

Haomin Chen

PhD of Computer Science

Johns Hopkins University

Biography

I am a PhD student of computer science in Johns Hopkins University. My research interests include transparent systems, medical imaging and computer vision. I have research experience about transparent deep learning systems for both 2D and 3D medical images; both radiological and pathological images; both classification and detection problems, with both PyTorch and Tensorflow.

Download my resumé.

Interests
  • Computer Vision
  • Medical Imaging
  • Transparent systems
Education
  • PhD in Computer Science, 2018-present

    Johns Hopkins University

  • M.A. in Statistics, 2016-2017

    Columbia University

  • BSc in Physics, 2012-2016

    Fudan Univerisity

Skills

Python
PyTorch
TensorFlow
Matlab
C++
R

Experience

 
 
 
 
 
PingAn
Applied Research Intern
PingAn
May 2019 – Dec 2019 Bethesda

Symmetric learning for Fracture Detection in Pelvic Trauma X-ray:

  • Accepted by ECCV with poster.
  • Achieved 98.8% AUC for all fractures, with is 0.8% higher than baseline model.
  • Created alignment between original/flipped images according to pelvic structure landmark detection.
  • Constructed Siamese network with fusion layer to incorporate symmetric information.
  • Aligned Siamese features instead of input images to reduce distortion artifacts.
  • Applied pixel-wise contrastive loss to learn pathologically asymmetric information explicitly.
 
 
 
 
 
NVIDIA
Applied Research Intern
NVIDIA
May 2018 – Dec 2018 Bethesda

Deep Hierarchical Multi-label Classification of Chest X-ray:

  • Accepted by MIDL 2019 with oral and special invitation to Journal Medical Image Analysis.
  • Achieved 89% AUC of all diseases on PLCO dataset, which is the state of the art.
  • Constructed Hierarchical label structure following clinical taxonomy.
  • Trained with conditional probabilities first and then fine-tuned with full probabilities training.
  • Derived a numerically stable formulation to calculate the cross entropy loss using full probabilities.
  • Introduced conditional AUCs for hierarchical-label performance evaluation.
 
 
 
 
 
PingAn
Applied Research Intern
PingAn
May 2017 – Aug 2017 Shanghai

Lung nodule detection in CT images:

  • Achieved rank 6 out of 2887 teams in the Skylake competition sponsored by Intel and Alibaba.
  • Applied PyTorch, 3D UNet and Caffe, Faster RCNN to detect lung nodules in 1000 CT scans.
  • Truncated the last two deconvolution layers in U Net changed the output as RPN structure and added new output branches before every deconvolution layer.
  • Used fusion method to achieve false positive reduction.

Recent Publications

(2020). Gene Expression Profile Prediction in Uveal Melanoma Using Deep Learning: A Pilot Study for the Development of an Alternative Survival Prediction Tool. In Ophthalmology Retina.

PDF Cite

(2019). Deep hierarchical multi-label classification of chest X-ray images. In MIDL.

PDF Cite

Contact