Challenge "Screening and Diagnosis of esophageal cancer" by Mauna Kea

The goal of this challenge is to build an image classifier to assist physicians in the screening and diagnosis of esophageal cancer. Such a tool would have a massive impact on patient management and patient lives.

There are 11161 images acquired from 61 patients to be classified as:

• Class 0: Squamous Epithelium
• Class 1: Intestinal Metaplasia
• Class 2: Gastric Metaplasia
• Class 3: Dysplasia/Cancer

The split between the training and the two test sets is 80%-10%-10%.

The training set is made of 9446 images, acquired from 44 patients:

• Class 0: 1469 images
• Class 1: 3177 images
• Class 2: 1206 images
• Class 3: 3594 images

The two test sets, Test_1 and Test_2, are as follow:

• 893 images acquired from 10 patients
• 822 images acquired from 7 patients

The total is 1715 images acquired from 17 patients.

We also know that if an image acquired from a patient is in the training set, there is no image acquired from the same patient in any test sets.

Non weighted Multiclass accuracy = (Number of images correctly classified) / (Number of images in the test set)

Goal: 99%

Baseline: 75% (simple CNN)

• Data preprocessing:

  • remove duplicates with dhash (9446 images -> 6973 images)
  • split 6973 left images in train/test/valid (80%/10%/10%)
  • No split by patient (I assumed it was not necessary since I had removed nearly identical images, maybe it's a bad idea...)

• Class imbalance:

  • over-sampling during training
  • Weighted CE loss for testing

• Training:

  • simple data aug (flips, random rotations)
  • resized to (224, 224), (299, 299) for InceptionResNetV2
  • Pretrained weights from imagenet
  • SGD
  • LR scheduler : cosine annealing + warm restarts (60 epochs with 3 periods of 20 epochs)
  • no regularization needed

• Pseudo-labeling:

  • on test set (1715 images)
  • predict labels for each image
  • if all (or more than X%) images that belong to the same patient have the same prectited class => add patient to train set

• Ensemble:

  • with probabilites from softmax

• Focal loss
• Adam/AdamW (converges fast but not as good as SGD)
• Adabound/Amsbound

• Other EfficientNet models
• TTA
• For pseudo-labeling:
  • https://arxiv.org/abs/1904.12848v2
  • https://arxiv.org/abs/1905.02249v1
• Hard mining
• ridge regression : http://blog.kaggle.com/2017/10/17/planet-understanding-the-amazon-from-space-1st-place-winners-interview/

Public Leaderboard	Final (private) Leaderboard