Oracle Imitating Selection

Framework Overview

Purpose

To develop more effective Active Learning (AL) patch recommendations

Methods

1. Building Oracle Selection Classification Model (Version 1)

   • Construct a simple classification model for predicting between two classes: oracle selection or not

2. Enhanced Model (Version 2)

   • Extract features from images
   • Concatenate image features with meta-information:
     • Whole Slide Image (WSI) ground truth
     • Patch prediction from DenseNet201 pretrained model (updated --> AL)
     • Patch confidence score from DenseNet201 pretrained model (updated --> AL)

3. Unified Model (Version 3)

   • Similar to Version 2, with the exception that the feature extractor and imitation prediction (simple model) are integrated into a single model
   • Validate the model to the imbalanced dataset (all)

Active Learning Strategy

Implements an AL strategy that combines the uncertainty from the AL process and the probability of oracle selection from the oracle-imitating selection model