Skip to content

Latest commit

 

History

History
167 lines (152 loc) · 7.83 KB

README.md

File metadata and controls

167 lines (152 loc) · 7.83 KB
Raw

SPIE Medical Imaging 2019

Sunday

Deep Learning Session 1

AI + Medical Imaging in China: Present and Future

  • Presenter Dr. Shiyuan Li from Changzheng Hospital in China
  • High quality medical services are needed
    • Aging population
  • Pain point of Medical Imaging in China
    • Inaccurate diagnosis
    • Scarcity of supply: 20% annual growth rate of data, 4.1% of radiologists
    • Low efficiency
  • AI + Medicine
    • Disease detection
    • Lesion quantification
    • Malignancy diagnosis
    • Therapeutic evaluation
  • Compared with US, China has way too many medical imaging AI company, but not enough new drug discovery, patient management, etc.
  • Usage:
    • lung nodule: daily use
    • fracture and ICH: ER
  • Iterative AI model
    • Refinement using hospital data
    • False positives gradually reduce to acceptable levels
  • AI tool for measuring intracerebral hemorrhage (ICH)
    • x ml --> y ml
  • Bone Age analysis
    • Parents height, child's height and weight, evaluates the status of the child's growth (too over-weight? ahead of the growth curve)
    • Shows the growth curve (height/weight vs age), and the child's position (a single point) in the curves
  • Large medical imaging database
    • National Database of ultrasound
    • National radiology database
  • CFDA
    • Class II: assistive diagnostic suggestion/recommendation, regulated by local provincial FDA
    • Class III: give definitive diagnostic suggestion/recommendation, regulated by CFDA

Impact of imprinted labels on deep learning classification of AP and PA thoracic radiographs

  • PA: Size of hard shadow, visual cues: scapula/clavicle positions
  • a CNN classifier to classify AP/PA
  • 2000 images, may contain foreign objects, 65:20:15 splits
  • Wording of imprinted labels: portable, AP, Erect, semi erect
  • 97% ROC AUC for images with labels, 95% ROC AUC for images without labels
  • Q&A:
    • Only DR is used
    • AP often contain more foreign objects
    • No CAM is used to visualize, this is the next step

Deep learning method for tumor segmentation in breast DCE-MRI

  • previous methods: fuzzy c-means (Zhang et al), watershed
  • 2D vs 3D methods: U-Net used
  • data: slices with tumors
  • 3D U-Net has fewer false positives
  • Q&A:
    • Test: how about false positives in healthy regions? ROI is specified by doctors. This is only for segmentation.
    • Patient split: yes

Deep Learning Session 2

Optic disc segmentation in fundus images

  • MWSSIDOR, open dataset, 1200 images
  • ROI abstracted from fundus images with RetinaNet
  • Segmentation done with U-Net

Learning cross-protocol radiomics and deep feature standardization from CT images of texture phantoms

  • Standardize features across scanners
  • Texture phantom scanned by various scanners
  • Classify scanned texture to learn features stable across scanners
  • Domain adversarial training with gradient reverse layer to forget about domain

A data interpretation approach for deep learning-based prediction models

  • Model interpretability
    • Radiomics vs deep learning
  • CNN is ruthless in finding discriminative features. Here is a cool story from the Pentagon.
  • Scheme 1: modify input images to examine CNN models wrt changes in ROI (occlusion test)
    • 4x4 grid: keep information inside one grid, but mask all other 15 girds (This does not quite make sense)
  • Scheme 2: CAM based methods
  • Notes: not very convincing

Breast III and Heart

Visual evidence for interpreting diagnostic decision of deep neural network in computer-aided diagnosis

  • Deep learning has limited interpretability
  • Solution: CAM (class activation map) or gradCAM
  • Inspired by Radiologists' interpretation: BI-RADS
  • Margin interpretation + shape interpretation
    • Interpretation loss
    • Consistency loss
    • Extra supervision loss
  • Note: the losses are too complicated to replicate

U-Net inspired architecture ensembles for left atrial segmentation

  • Ensemble of 15 different U-Net based on different feature extractors (ResNet, DenseNet, SENet, etc)
  • Some individual model has wide error bars
  • The ensemble model does not have clear efficiency over some of the good performing models

Monday

Deep Learning

Large-scale evaluation of multiresolution V-Net for organ segmentation in image guided radiation therapy

  • From United Imaging, trtment planning
  • Coarse to fine. Downsample, use v-net to get mask, then crop patches, feed into v-net, segment
  • Model compression and memory consumption
    • V-net too large for production.
      • Reduce kernel size (5x5x5 --> 3x3x3): 250 MB --> 57 MB
      • Use bottleneck to replace conv layers with large channel sizes. Output_ch x input_ch x K x K x K ~ num of channels. C x C x 3 --> C/N x C x 1 x 1 x 1 --> C/N x C/N x 3 x 3 x 3 --> C x C/N x 1 x 1 x 1. N = 4 --> a factor of 12 compression ratio. 250 MB --> 57 MB --> 8.8 MB.
    • GPU memory use: downsample image by 2
  • GTX 1080 card, 0.7 second
  • Demo: Atlas based: 3 min, V-Net: 7 seconds. More accurate than atlas based methods, e.g. in liver tips.
  • United Imaging used highly optimized backend (cuda lib) for inference. Pytorch and TF use too much memory for inference.
  • Dice loss is better than cross entropy
    • Dice loss does not lead to overfitting
    • Cross entropy (weighted version or focal loss) often overfits and requires validation set to pick a model. This is not always possible as the validation set is too large to evaluate after every epoch.

StreoScenNet: surgical stereo robotic scene segmentation

  • This talk is very interesting.
  • End to end training of a single network for all tasks
  • multi-encoder and single decoder
  • Each encoder initialized with imagenet weights.
    • binary segmentation
    • instrument segmentation
    • part segmentation
  • The results of each encoder is summed, concatenated with the same level of decoder, and generating 3 masks.
  • Data: MICCAI 2917 endoscopic challenge
    • Left and right. only left are annotated
  • Q&A:
    • enforce consistency? no. the author does not think that would matter much.
  • This is very similar to the Y-Net.

Deep learning based 2.5D flow field estimation for maximum intensity projections of 4D optical coherence tomography

  • OCT: cloud points?
  • Flat 3D to 2D with MIP along depth
  • census transform
  • Overall approach:
    • Approximate x, y flow, then align x, y and concate and estimate the z-flow.
  • Encoder-decoder: ERFnet
  • Multi-loss, but the strong supervised loss is not needed for convergence.
  • Level-wise training is critical for convegence

Automatic vertebrae localization in spine CT: a deep-learning approach for image guidance and surgical data science

  • From Jeffrey Siewerdsen's group
  • Challenge: variable imaging protocol, poor image quality, foreign objects
  • YOLO to regress points. Essentially get rid of width and height loss terms.
  • 2D vs 3D:
    • 2D on each slice, but use aggregation network to combine the feature maps to regress x, y, z centroid
    • 2D takes significantly less resource
    • Not very stable
  • Deeper network, use imagenet pre-train
    • Use detection sagittal slices and coronal slices
  • Faster RCNN ortho 2D gives best results
  • GT and prediction are in 3D, visualization are in 2D
  • Used closest GT for evaluation