/PANDA-Prostate-cANcer-graDe-Assessment

Primary LanguageJupyter Notebook

PANDA-Prostate-cANcer-graDe-Assessment

The challenge in this Kaggle competition is to classify the severity of prostate cancer (Gleason Scores/ISUP Grades) from microscopy scans of prostate biopsy samples.

Example Biopsy Sample and the grading system

Here are a few approaches I tried to approach the problem with brief description and achieved score on public leader board (Score is Quadratic Weighted Kappa (QWK) in this competition).

Experiment #1 (0.47 on public LB)

  • Used model: DenseNet121, trained from scratch
  • No image augmentations
  • Other settings:
    • Image size: 256 x 256 size patches of original image
    • No CV
    • 20 epochs
    • Batch size: 16 images

Experiment #2 (0.51 on public LB)

  • Used model: DenseNet121, pre-trained ImageNet weights
  • Basic shift, scale, rotation and flips
  • Other settings:
    • Image size: 256 x 256 size patches of original image
    • No CV
    • 25 epochs
    • Batch size: 16 images

Experiment #3 (0.59 on public LB)

  • Used model: DenseNet121, pre-trained ImageNet weights
  • Basic shift, scale, rotation and flips
  • Other settings:
    • Image size: 256 x 256 size patches of original image
    • 5 Fold CV
    • 25 epochs
    • Batch size: 16 images

Experiment #4 (0.51 on public LB)

  • Used model: DenseNet121, pre-trained ImageNet weights
  • Basic shift, scale, rotation and flips
  • Other settings:
    • Trained on Gleason score instead of directly on ISUP grades
    • Image size: 256 x 256 size patches of original image
    • No CV
    • 25 epochs
    • Batch size: 16 images

Experiment #5 (0.60 on public LB)

  • Used model: DenseNet121, pre-trained ImageNet weights
  • Basic shift, scale, rotation and flips
  • Other settings:
    • Image size: 256 x 256 size patches of original image
    • 5 Fold CV
    • 25 epochs
    • Batch size: 16 images
    • TTA (Test Time Augmentation)

Experiment #6 (0.55 on public LB)

  • Used model: DenseNet121, pre-trained ImageNet weights
  • Basic shift, scale, rotation and flips
  • Other settings:
    • Image size: 256 x 256 size patches of original image
    • No CV
    • 25 epochs
    • Batch size: 16 images
    • Label Smoothing

Experiment #7 (0.56 on public LB)

  • Used model: DenseNet121, pre-trained ImageNet weights
  • Basic shift, scale, rotation and flips
  • Other settings:
    • Image size: 256 x 256 size patches of original image
    • No CV
    • 25 epochs
    • Batch size: 16 images
    • Label Smoothing
    • TTA

Experiment #8 (0.53 on public LB)

  • Used model: DenseNet121, pre-trained ImageNet weights
  • Basic shift, scale, rotation and flips
  • Other settings:
    • Image size: 256 x 256 size patches of original image
    • 5-fold CV
    • 30 epochs
    • Batch size: 16 images
    • Label Smoothing
    • TTA