Intelligent Systems for Medical Image Processing (Radboud Universiteit, Netherlands)

18 flashcards · Computer Science · by Locke Lamora

  1. Sensitivity

    TPTP+FN\frac{TP}{TP+FN}TP+FNTP​

  2. Specificity

    TNTN+FP\frac{TN}{TN+FP}TN+FPTN​

  3. Accuracy

    TP+TNTP+TN+FP+FN\frac{TP+TN}{TP+TN+FP+FN}TP+TN+FP+FNTP+TN​

  4. Dice Coefficient

    2TP2TP+FN+FP=2∣R∩S∣∣R∣+∣S∣\frac{2TP}{2TP+FN+FP}=\frac{2|R\cap S|}{|R|+|S|}2TP+FN+FP2TP​=∣R∣+∣S∣2∣R∩S∣​

  5. ROC

    x-axis: 1-specificity/FPR y-axis: TPR/sensitivity

  6. Precision

    TPTP+FP\frac{TP}{TP+FP}TP+FPTP​

  7. Segmentation Performance Metrics:

    Jaccard Dice Score Accuracy

  8. Classification Performance Metrics:

    ROC AUC

  9. Detection Performance Metrics:

    Precision F1-Score FROC

  10. FROC

    x-axis: average number of false positives y-axis: TPR/sensitivity

  11. Convolutions

    Valid convolution: no padding Same convolution: padding

  12. Outputsize formulas:

    o=n−(f−1)o=n-(f-1)o=n−(f−1) o=n−(f−1)(d−1)−(f−1)o=n-(f-1)(d-1)-(f-1)o=n−(f−1)(d−1)−(f−1) o=(n+2p−(f−1)(d−1)−f)/s+1o=(n+2p-(f-1)(d-1)-f)/s + 1o=(n+2p−(f−1)(d−1)−f)/s+1

  13. Image Registration

    Find same thing in similar picture (using transformation) Pipeline: Center alignment Translation Affine registration Deformable registration

  14. Popular ConvNets for Classification:

    AlexNet (ReLU + Dropout) VVGnet (19 Layers, fixed 3x3 filters) GoogleNet (22 Layers) ResNet (residual blocks for gradient flow) DenseNet (has residual blocks to all past layers)

  15. Hard negative mining

    Train a second network on the hard negatives only

  16. R-CNN

    R-CNN: region proposal by selective search regions resized and processed using AlexNet SVM gives class Fast R-CNN: Process image first, then crop Faster R-CNN: One network for everything

  17. YOLO

    Divide image in grid each cell predicts bounding box and confidence score if cell is center, it is responsible for detecting that object Confidence = probability of object * IoU Each grid cell predict conditional class probabilities Output shape is SxSx(B*5+C) has custom loss function for location, height and width of bounding boxes and confidence Limitations: each grid cell can predict only two boxes and one class, limiting the number of nearby object that can be predicted struggles with small…

  18. Otsu algorithm

    Tries multiple thresholds for image segmentation Minimizes intra-class variance